{ "domain": "darkgpt.org", "systemName": "DarkGPT.org Uncensored LLMs Ranking System", "snapshotDate": "2026-11-01", "syncedLiveRepos": 1695, "totalIndexedModels": 450, "models": [ { "id": "dphn-dolphin-2-9-1-yi-1-5-34b", "name": "dolphin-2.9.1-yi-1.5-34b", "hfRepo": "dphn/dolphin-2.9.1-yi-1.5-34b", "creator": "dphn", "baseArchitecture": "01.AI Yi (34B)", "paramsBillions": 34, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-18", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 99, "mmluPro": 65.3, "eqBench": 83.4, "ifeval": 82.2, "hfDownloads30d": 4778759, "hfLikes": 90, "ollamaCommand": "dphn/dolphin-2.9.1-yi-1.5-34b", "vramQ4GB": 20, "vramQ8GB": 38, "vramFP16GB": 74, "summary": "Uncensored DPO/SFT fine-tune of 01.AI Yi (34B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~20 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 34B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-8-27b-abliterated-gguf", "name": "Huihui-Qwen3.8-27B-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-16", "lastVerified": "2026-10-10", "ugiScore": 73.6, "complianceRate": 98.6, "mmluPro": 67, "eqBench": 81.6, "ifeval": 84.5, "hfDownloads30d": 3157401, "hfLikes": 1622, "ollamaCommand": "huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "jonathancoletti-qwen3-8-27b-uncensored-gguf", "name": "Qwen3.8-27B-Uncensored-GGUF", "hfRepo": "JonathanColetti/Qwen3.8-27B-Uncensored-GGUF", "creator": "JonathanColetti", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-14", "lastVerified": "2026-11-01", "ugiScore": 74.3, "complianceRate": 98.5, "mmluPro": 66.2, "eqBench": 82.2, "ifeval": 84.3, "hfDownloads30d": 2598773, "hfLikes": 1840, "ollamaCommand": "JonathanColetti/Qwen3.8-27B-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by JonathanColetti, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-8-27b-uncensored-hauhaucs-aggressive-mtp-gguf", "name": "Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF", "hfRepo": "HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-17", "lastVerified": "2026-11-01", "ugiScore": 73.6, "complianceRate": 99.4, "mmluPro": 67.8, "eqBench": 82.8, "ifeval": 83.6, "hfDownloads30d": 2380949, "hfLikes": 2186, "ollamaCommand": "HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-6-35b-a3b-abliterated-mtp-gguf", "name": "Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF", "hfRepo": "huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-05-19", "lastVerified": "2026-05-19", "ugiScore": 73.4, "complianceRate": 99, "mmluPro": 69.2, "eqBench": 83.1, "ifeval": 84.3, "hfDownloads30d": 2279692, "hfLikes": 69, "ollamaCommand": "huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davidau-qwen3-8-27b-turbo-fable-cold-fusion-735-882-heretic-uncensored-neo-coder-max-mtp-gguf", "name": "Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF", "hfRepo": "DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF", "creator": "DavidAU", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-01", "lastVerified": "2026-11-01", "ugiScore": 73.5, "complianceRate": 99.1, "mmluPro": 70.5, "eqBench": 80, "ifeval": 86.6, "hfDownloads30d": 1451434, "hfLikes": 2280, "ollamaCommand": "DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-6-35b-a3b-uncensored-hauhaucs-aggressive", "name": "Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-04-17", "lastVerified": "2026-11-01", "ugiScore": 75.2, "complianceRate": 99, "mmluPro": 68.6, "eqBench": 83.4, "ifeval": 85.2, "hfDownloads30d": 1106939, "hfLikes": 4069, "ollamaCommand": "HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.6 (35B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "0bserverx-qwen3-8-27b-heretic-abliterated-uncensored-gguf", "name": "Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "hfRepo": "0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "creator": "0bserverx", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-14", "lastVerified": "2026-11-01", "ugiScore": 73.7, "complianceRate": 98.1, "mmluPro": 68, "eqBench": 81.8, "ifeval": 84.6, "hfDownloads30d": 1149119, "hfLikes": 814, "ollamaCommand": "0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by 0bserverx, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "hauhaucs-gemma-4-e4b-uncensored-hauhaucs-aggressive", "name": "Gemma-4-E4B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Gemma 2 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-04-02", "lastVerified": "2026-11-01", "ugiScore": 63.3, "complianceRate": 99.4, "mmluPro": 51.3, "eqBench": 73.2, "ifeval": 72.6, "hfDownloads30d": 831873, "hfLikes": 1514, "ollamaCommand": "HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (4B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~3.4 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 4B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "jegreene-swift-1-5-qwen3-8-27b-heretic-gsq-rco-gguf", "name": "Swift-1.5-Qwen3.8-27B-heretic-GSQ-RCO-GGUF", "hfRepo": "jegreene/Swift-1.5-Qwen3.8-27B-heretic-GSQ-RCO-GGUF", "creator": "jegreene", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-29", "lastVerified": "2026-11-01", "ugiScore": 73, "complianceRate": 99.2, "mmluPro": 68, "eqBench": 81, "ifeval": 83.9, "hfDownloads30d": 769010, "hfLikes": 27, "ollamaCommand": "jegreene/Swift-1.5-Qwen3.8-27B-heretic-GSQ-RCO-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by jegreene, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "hauhaucs-gemma4-12b-qat-uncensored-hauhaucs-balanced", "name": "Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced", "hfRepo": "HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced", "creator": "HauhauCS", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-06-22", "lastVerified": "2026-11-01", "ugiScore": 68.1, "complianceRate": 99.2, "mmluPro": 58.5, "eqBench": 77.5, "ifeval": 77.5, "hfDownloads30d": 714257, "hfLikes": 505, "ollamaCommand": "HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (12B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-5-9b-uncensored-hauhaucs-aggressive", "name": "Qwen3.5-9B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-03-04", "lastVerified": "2026-11-01", "ugiScore": 69.5, "complianceRate": 98.7, "mmluPro": 59, "eqBench": 76.4, "ifeval": 79.4, "hfDownloads30d": 662378, "hfLikes": 2420, "ollamaCommand": "HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.5 (9B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~6.2 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 9B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "bahushruth-qwen3-6-35b-a3b-abliterated-v4", "name": "Qwen3.6-35B-A3B-abliterated-v4", "hfRepo": "Bahushruth/Qwen3.6-35B-A3B-abliterated-v4", "creator": "Bahushruth", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-11", "lastVerified": "2026-11-01", "ugiScore": 74.5, "complianceRate": 98.9, "mmluPro": 69, "eqBench": 83.6, "ifeval": 85.4, "hfDownloads30d": 579362, "hfLikes": 13, "ollamaCommand": "Bahushruth/Qwen3.6-35B-A3B-abliterated-v4", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by Bahushruth, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "philbert440-qwen3-8-27b-uncensored-cyber-gguf", "name": "Qwen3.8-27B-Uncensored-Cyber-GGUF", "hfRepo": "philbert440/Qwen3.8-27B-Uncensored-Cyber-GGUF", "creator": "philbert440", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-17", "lastVerified": "2026-11-01", "ugiScore": 73, "complianceRate": 99.6, "mmluPro": 66.8, "eqBench": 82.3, "ifeval": 83, "hfDownloads30d": 538634, "hfLikes": 76, "ollamaCommand": "philbert440/Qwen3.8-27B-Uncensored-Cyber-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by philbert440, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "davidau-qwen3-6-27b-fable-fusion-711-uncensored-heretic-nm-dau-neo-max-mtp-gguf", "name": "Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF", "hfRepo": "DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF", "creator": "DavidAU", "baseArchitecture": "Qwen 3.6 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-17", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 98.9, "mmluPro": 68.9, "eqBench": 81.1, "ifeval": 83.6, "hfDownloads30d": 432769, "hfLikes": 2487, "ollamaCommand": "DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.6 (27B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "luffythefox-qwen3-6-35b-a3b-uncensored-genesis-hermes-gguf", "name": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "hfRepo": "LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "creator": "LuffyTheFox", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-07-13", "lastVerified": "2026-11-01", "ugiScore": 74.8, "complianceRate": 98.6, "mmluPro": 69.7, "eqBench": 83.8, "ifeval": 84.8, "hfDownloads30d": 455154, "hfLikes": 721, "ollamaCommand": "LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.6 (35B) by LuffyTheFox, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-3-llama-3-1-8b", "name": "Hermes-3-Llama-3.1-8B", "hfRepo": "NousResearch/Hermes-3-Llama-3.1-8B", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-07-28", "lastVerified": "2024-09-08", "ugiScore": 67.5, "complianceRate": 98.1, "mmluPro": 54.9, "eqBench": 78.8, "ifeval": 78.1, "hfDownloads30d": 457699, "hfLikes": 541, "ollamaCommand": "NousResearch/Hermes-3-Llama-3.1-8B", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-gemma4-26b-a4b-qat-uncensored-hauhaucs-balanced-mtp", "name": "Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP", "hfRepo": "HauhauCS/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP", "creator": "HauhauCS", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-06-24", "lastVerified": "2026-11-01", "ugiScore": 71, "complianceRate": 99.3, "mmluPro": 64.2, "eqBench": 80.7, "ifeval": 80.4, "hfDownloads30d": 393086, "hfLikes": 337, "ollamaCommand": "HauhauCS/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (26B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 26B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "soyaakinohara-spark-x2-5-4b-heretic-jp-gguf", "name": "Spark-X2.5-4B-Heretic-jp-gguf", "hfRepo": "soyaakinohara/Spark-X2.5-4B-Heretic-jp-gguf", "creator": "soyaakinohara", "baseArchitecture": "Open Transformer (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-04", "lastVerified": "2026-11-01", "ugiScore": 62.3, "complianceRate": 97.6, "mmluPro": 52.8, "eqBench": 74.1, "ifeval": 72.5, "hfDownloads30d": 383365, "hfLikes": 17, "ollamaCommand": "soyaakinohara/Spark-X2.5-4B-Heretic-jp-gguf", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Open Transformer (4B) checkpoint by soyaakinohara, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "failspy-meta-llama-3-8b-instruct-abliterated-v3", "name": "Meta-Llama-3-8B-Instruct-abliterated-v3", "hfRepo": "failspy/Meta-Llama-3-8B-Instruct-abliterated-v3", "creator": "failspy", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-20", "lastVerified": "2024-05-30", "ugiScore": 66.4, "complianceRate": 97.9, "mmluPro": 55.8, "eqBench": 76.1, "ifeval": 76.6, "hfDownloads30d": 347448, "hfLikes": 69, "ollamaCommand": "failspy/Meta-Llama-3-8B-Instruct-abliterated-v3", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3 (8B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "atomicchat-lfm2-5-24b-a2b-uncensored-gguf", "name": "LFM2.5-24B-A2B-Uncensored-GGUF", "hfRepo": "AtomicChat/LFM2.5-24B-A2B-Uncensored-GGUF", "creator": "AtomicChat", "baseArchitecture": "Open Transformer (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-10-09", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.8, "mmluPro": 62.2, "eqBench": 80.9, "ifeval": 82, "hfDownloads30d": 309727, "hfLikes": 72, "ollamaCommand": "AtomicChat/LFM2.5-24B-A2B-Uncensored-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (24B) by AtomicChat, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 24B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-qwen2-5-vl-7b-instruct-abliterated", "name": "Qwen2.5-VL-7B-Instruct-abliterated", "hfRepo": "huihui-ai/Qwen2.5-VL-7B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-02-17", "lastVerified": "2026-11-01", "ugiScore": 67.4, "complianceRate": 97.9, "mmluPro": 58.8, "eqBench": 75.3, "ifeval": 78.1, "hfDownloads30d": 298926, "hfLikes": 64, "ollamaCommand": "huihui-ai/Qwen2.5-VL-7B-Instruct-abliterated", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 (7B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "ttindall-qwen3-8-27b-defiance-uncensored-mtp-gguf", "name": "Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF", "hfRepo": "ttindall/Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF", "creator": "ttindall", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-10-17", "lastVerified": "2026-11-01", "ugiScore": 73.4, "complianceRate": 99.2, "mmluPro": 67.6, "eqBench": 82.6, "ifeval": 83.4, "hfDownloads30d": 294520, "hfLikes": 205, "ollamaCommand": "ttindall/Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by ttindall, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-5-4b-uncensored-hauhaucs-aggressive", "name": "Qwen3.5-4B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.5 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-03-03", "lastVerified": "2026-11-01", "ugiScore": 65.7, "complianceRate": 99.4, "mmluPro": 54.4, "eqBench": 73.6, "ifeval": 75.5, "hfDownloads30d": 243837, "hfLikes": 707, "ollamaCommand": "HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.5 (4B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~3.4 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 4B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "atomicchat-minicpm-v-4-7-35b-a3b-uncensored-gguf", "name": "MiniCPM-V-4.7-35B-A3B-Uncensored-GGUF", "hfRepo": "AtomicChat/MiniCPM-V-4.7-35B-A3B-Uncensored-GGUF", "creator": "AtomicChat", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-10-12", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 98.1, "mmluPro": 64.8, "eqBench": 83.7, "ifeval": 82.2, "hfDownloads30d": 174737, "hfLikes": 44, "ollamaCommand": "AtomicChat/MiniCPM-V-4.7-35B-A3B-Uncensored-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (35B) by AtomicChat, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-5-122b-a10b-uncensored-hauhaucs-aggressive", "name": "Qwen3.5-122B-A10B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.5-122B-A10B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.5 (122B)", "paramsBillions": 122, "parameterClass": "Frontier (100B+)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-03-22", "lastVerified": "2026-11-01", "ugiScore": 77.5, "complianceRate": 98.8, "mmluPro": 73.6, "eqBench": 85.4, "ifeval": 87.6, "hfDownloads30d": 170126, "hfLikes": 192, "ollamaCommand": "HauhauCS/Qwen3.5-122B-A10B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 69, "vramQ8GB": 134, "vramFP16GB": 262, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.5 (122B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on multi-GPU workstations or cloud nodes (~69 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (69 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "azukivc-qwen3-8-27b-uncensored-gguf", "name": "Qwen3.8-27B-Uncensored-GGUF", "hfRepo": "azukivc/Qwen3.8-27B-Uncensored-GGUF", "creator": "azukivc", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-22", "lastVerified": "2026-11-01", "ugiScore": 72.9, "complianceRate": 98.7, "mmluPro": 66.7, "eqBench": 82.1, "ifeval": 82.9, "hfDownloads30d": 163559, "hfLikes": 3, "ollamaCommand": "azukivc/Qwen3.8-27B-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by azukivc, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "kujira-underdog-saluki-27b-1-0-mtp-abliterated-gguf", "name": "Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF", "hfRepo": "Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF", "creator": "Kujira", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-09", "lastVerified": "2026-11-01", "ugiScore": 72, "complianceRate": 98.6, "mmluPro": 65, "eqBench": 81.1, "ifeval": 82.4, "hfDownloads30d": 161171, "hfLikes": 36, "ollamaCommand": "Kujira/Underdog-Saluki-27B-1.0-MTP-Abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by Kujira, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davidau-qwen3-8b-64k-context-2x-josiefied-uncensored", "name": "Qwen3-8B-64k-Context-2X-Josiefied-Uncensored", "hfRepo": "DavidAU/Qwen3-8B-64k-Context-2X-Josiefied-Uncensored", "creator": "DavidAU", "baseArchitecture": "Qwen 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-05-06", "lastVerified": "2026-11-01", "ugiScore": 67.6, "complianceRate": 98.9, "mmluPro": 59.5, "eqBench": 77.6, "ifeval": 77.4, "hfDownloads30d": 147740, "hfLikes": 22, "ollamaCommand": "DavidAU/Qwen3-8B-64k-Context-2X-Josiefied-Uncensored", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3 (8B) by DavidAU, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-5-35b-a3b-uncensored-hauhaucs-aggressive", "name": "Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.5 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-03-10", "lastVerified": "2026-11-01", "ugiScore": 74.7, "complianceRate": 99.3, "mmluPro": 68, "eqBench": 83.6, "ifeval": 84.7, "hfDownloads30d": 119744, "hfLikes": 1537, "ollamaCommand": "HauhauCS/Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.5 (35B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "os-software-ternary-bonsai-2-27b-uncensored-heretic-gguf", "name": "Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF", "hfRepo": "OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF", "creator": "OS-Software", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-18", "lastVerified": "2026-11-01", "ugiScore": 71.5, "complianceRate": 98.1, "mmluPro": 64.1, "eqBench": 81.1, "ifeval": 81.9, "hfDownloads30d": 137214, "hfLikes": 147, "ollamaCommand": "OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by OS-Software, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "orcarouter-qwen3-8-27b-uncensored-gguf", "name": "Qwen3.8-27B-Uncensored-GGUF", "hfRepo": "orcarouter/Qwen3.8-27B-Uncensored-GGUF", "creator": "orcarouter", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-16", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 98.2, "mmluPro": 66.3, "eqBench": 82.2, "ifeval": 82.4, "hfDownloads30d": 104353, "hfLikes": 1656, "ollamaCommand": "orcarouter/Qwen3.8-27B-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by orcarouter, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "zaakirio-gemma-4-12b-it-uncensored-gguf", "name": "gemma-4-12b-it-uncensored-GGUF", "hfRepo": "zaakirio/gemma-4-12b-it-uncensored-GGUF", "creator": "zaakirio", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-06-04", "lastVerified": "2026-11-01", "ugiScore": 68.7, "complianceRate": 98.2, "mmluPro": 57.7, "eqBench": 77.8, "ifeval": 78.1, "hfDownloads30d": 104593, "hfLikes": 73, "ollamaCommand": "zaakirio/gemma-4-12b-it-uncensored-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (12B) by zaakirio, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "wangzhang-gemma-4-31b-it-abliterated", "name": "gemma-4-31B-it-abliterated", "hfRepo": "wangzhang/gemma-4-31B-it-abliterated", "creator": "wangzhang", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-05", "lastVerified": "2026-11-01", "ugiScore": 72.1, "complianceRate": 98.6, "mmluPro": 65.8, "eqBench": 81.5, "ifeval": 82.5, "hfDownloads30d": 101773, "hfLikes": 43, "ollamaCommand": "wangzhang/gemma-4-31B-it-abliterated", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by wangzhang, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "culturerevolt-gemma-4-12b-heretic-abliterated-gguf", "name": "gemma-4-12b-heretic-abliterated-GGUF", "hfRepo": "culturerevolt/gemma-4-12b-heretic-abliterated-GGUF", "creator": "culturerevolt", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-05", "lastVerified": "2026-11-01", "ugiScore": 67.6, "complianceRate": 99.1, "mmluPro": 59.8, "eqBench": 79.3, "ifeval": 77.9, "hfDownloads30d": 98204, "hfLikes": 50, "ollamaCommand": "culturerevolt/gemma-4-12b-heretic-abliterated-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 2 (12B) checkpoint by culturerevolt, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "isvalorum-qwen3-8-35b-a3b-distill-mtp-apex-i-miniplus-v2-1-abliterated-gguf", "name": "Qwen3.8-35B-A3B-Distill-MTP-APEX-I-MiniPlus-V2.1-Abliterated-GGUF", "hfRepo": "IsValorum/Qwen3.8-35B-A3B-Distill-MTP-APEX-I-MiniPlus-V2.1-Abliterated-GGUF", "creator": "IsValorum", "baseArchitecture": "Qwen 3.8 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-20", "lastVerified": "2026-11-01", "ugiScore": 74.5, "complianceRate": 97.7, "mmluPro": 68.8, "eqBench": 83.8, "ifeval": 85.4, "hfDownloads30d": 96973, "hfLikes": 54, "ollamaCommand": "IsValorum/Qwen3.8-35B-A3B-Distill-MTP-APEX-I-MiniPlus-V2.1-Abliterated-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.8 (35B) checkpoint by IsValorum, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v4-3-gguf", "name": "Cydonia-24B-v4.3-GGUF", "hfRepo": "TheDrummer/Cydonia-24B-v4.3-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-11-08", "lastVerified": "2025-12-17", "ugiScore": 70, "complianceRate": 98.7, "mmluPro": 61.8, "eqBench": 85.1, "ifeval": 80.4, "hfDownloads30d": 95465, "hfLikes": 103, "ollamaCommand": "TheDrummer/Cydonia-24B-v4.3-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "yuyu1015-huihui-qwen3-30b-a3b-instruct-2507-abliterated-nvfp4", "name": "Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated-NVFP4", "hfRepo": "YuYu1015/Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated-NVFP4", "creator": "YuYu1015", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-12", "lastVerified": "2026-11-01", "ugiScore": 72.9, "complianceRate": 99.2, "mmluPro": 68.5, "eqBench": 83.1, "ifeval": 83.8, "hfDownloads30d": 95937, "hfLikes": 1, "ollamaCommand": "YuYu1015/Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated-NVFP4", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by YuYu1015, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "hauhaucs-gemma4-26b-a4b-uncensored-hauhaucs-balanced", "name": "Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced", "hfRepo": "HauhauCS/Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced", "creator": "HauhauCS", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-05-14", "lastVerified": "2026-11-01", "ugiScore": 70.9, "complianceRate": 98.4, "mmluPro": 63.6, "eqBench": 82.5, "ifeval": 80.3, "hfDownloads30d": 90792, "hfLikes": 308, "ollamaCommand": "HauhauCS/Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (26B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 26B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-gemma-4-e2b-uncensored-hauhaucs-aggressive", "name": "Gemma-4-E2B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Gemma 2 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-04-02", "lastVerified": "2026-11-01", "ugiScore": 61, "complianceRate": 99.4, "mmluPro": 47.1, "eqBench": 71.4, "ifeval": 70.1, "hfDownloads30d": 88919, "hfLikes": 361, "ollamaCommand": "HauhauCS/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (2B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.4 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 2B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "aeon-7-qwen3-6-35b-a3b-heretic-nvfp4", "name": "Qwen3.6-35B-A3B-heretic-NVFP4", "hfRepo": "AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4", "creator": "AEON-7", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-17", "lastVerified": "2026-11-01", "ugiScore": 74.3, "complianceRate": 98.1, "mmluPro": 68.8, "eqBench": 83.7, "ifeval": 85.2, "hfDownloads30d": 91917, "hfLikes": 74, "ollamaCommand": "AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by AEON-7, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "ajgazin-swift-1-5-qwen3-8-27b-uncensored-dynamic-mtp-gguf", "name": "Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF", "hfRepo": "ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF", "creator": "ajgazin", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-09-25", "lastVerified": "2026-11-01", "ugiScore": 73.1, "complianceRate": 98.1, "mmluPro": 66.3, "eqBench": 82.2, "ifeval": 83.1, "hfDownloads30d": 90409, "hfLikes": 56, "ollamaCommand": "ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by ajgazin, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-gemma4-31b-qat-uncensored-hauhaucs-balanced-mtp", "name": "Gemma4-31B-QAT-Uncensored-HauhauCS-Balanced-MTP", "hfRepo": "HauhauCS/Gemma4-31B-QAT-Uncensored-HauhauCS-Balanced-MTP", "creator": "HauhauCS", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Uncensored DPO / SFT", "releaseDate": "2026-06-24", "lastVerified": "2026-11-01", "ugiScore": 71.4, "complianceRate": 98.1, "mmluPro": 64.9, "eqBench": 82.9, "ifeval": 80.9, "hfDownloads30d": 86712, "hfLikes": 201, "ollamaCommand": "HauhauCS/Gemma4-31B-QAT-Uncensored-HauhauCS-Balanced-MTP", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored DPO/SFT fine-tune of Gemma 2 (31B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 31B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-pro-mistral-7b-gguf", "name": "Hermes-2-Pro-Mistral-7B-GGUF", "hfRepo": "NousResearch/Hermes-2-Pro-Mistral-7B-GGUF", "creator": "NousResearch", "baseArchitecture": "Mistral (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-03-02", "lastVerified": "2024-03-28", "ugiScore": 65.6, "complianceRate": 99.1, "mmluPro": 55, "eqBench": 76.2, "ifeval": 74.9, "hfDownloads30d": 78716, "hfLikes": 262, "ollamaCommand": "NousResearch/Hermes-2-Pro-Mistral-7B-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Mistral (7B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "lukey03-qwen3-5-9b-abliterated-gguf", "name": "Qwen3.5-9B-abliterated-GGUF", "hfRepo": "lukey03/Qwen3.5-9B-abliterated-GGUF", "creator": "lukey03", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-03", "lastVerified": "2026-11-01", "ugiScore": 67.4, "complianceRate": 99.2, "mmluPro": 61, "eqBench": 77.4, "ifeval": 78.2, "hfDownloads30d": 80979, "hfLikes": 99, "ollamaCommand": "lukey03/Qwen3.5-9B-abliterated-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by lukey03, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "fyb1214-huihui-qwen3-8-27b-abliterated-swift-1-5-gsq-rco-ninfer", "name": "Huihui-Qwen3.8-27B-Abliterated-Swift-1.5-GSQ-RCO-NInfer", "hfRepo": "fyb1214/Huihui-Qwen3.8-27B-Abliterated-Swift-1.5-GSQ-RCO-NInfer", "creator": "fyb1214", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-04", "lastVerified": "2026-11-01", "ugiScore": 73.7, "complianceRate": 98.7, "mmluPro": 68.4, "eqBench": 81.7, "ifeval": 84.6, "hfDownloads30d": 74179, "hfLikes": 10, "ollamaCommand": "fyb1214/Huihui-Qwen3.8-27B-Abliterated-Swift-1.5-GSQ-RCO-NInfer", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by fyb1214, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "ravichandranj-dolphin3-cyber-8b-gguf", "name": "Dolphin3-Cyber-8B-GGUF", "hfRepo": "RavichandranJ/Dolphin3-Cyber-8B-GGUF", "creator": "RavichandranJ", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-02-13", "lastVerified": "2026-11-01", "ugiScore": 66.7, "complianceRate": 99.6, "mmluPro": 56.7, "eqBench": 76.1, "ifeval": 76, "hfDownloads30d": 67507, "hfLikes": 286, "ollamaCommand": "RavichandranJ/Dolphin3-Cyber-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (8B) by RavichandranJ, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-6-27b-uncensored-hauhaucs-aggressive", "name": "Qwen3.6-27B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.6 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-04-22", "lastVerified": "2026-11-01", "ugiScore": 72.6, "complianceRate": 99.2, "mmluPro": 66.5, "eqBench": 82.1, "ifeval": 82.6, "hfDownloads30d": 59680, "hfLikes": 744, "ollamaCommand": "HauhauCS/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.6 (27B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nerkyor-qwen3-8-27b-efficientthink-uncensored-k3-opus5-grok4-6-gpt5-6sol-sft-simpo-dflash2-gguf", "name": "Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2-GGUF", "hfRepo": "nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2-GGUF", "creator": "nerkyor", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-09-03", "lastVerified": "2026-11-01", "ugiScore": 73.9, "complianceRate": 98.1, "mmluPro": 66.8, "eqBench": 81.3, "ifeval": 83.9, "hfDownloads30d": 67781, "hfLikes": 118, "ollamaCommand": "nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.8 (27B) by nerkyor, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-cyberstrike-offsec-35b-abliterated-gguf", "name": "Huihui-CyberStrike-OffSec-35B-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-CyberStrike-OffSec-35B-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-11", "lastVerified": "2026-08-12", "ugiScore": 71.7, "complianceRate": 97.7, "mmluPro": 67.4, "eqBench": 83.7, "ifeval": 82.1, "hfDownloads30d": 67808, "hfLikes": 56, "ollamaCommand": "huihui-ai/Huihui-CyberStrike-OffSec-35B-abliterated-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Open Transformer (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "0bserverx-qwen3-8-27b-heretic-gsq-rco-gguf", "name": "Qwen3.8-27B-Heretic-GSQ-RCO-GGUF", "hfRepo": "0bserverx/Qwen3.8-27B-Heretic-GSQ-RCO-GGUF", "creator": "0bserverx", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-20", "lastVerified": "2026-11-01", "ugiScore": 73.1, "complianceRate": 98.7, "mmluPro": 68.8, "eqBench": 82, "ifeval": 84, "hfDownloads30d": 64170, "hfLikes": 66, "ollamaCommand": "0bserverx/Qwen3.8-27B-Heretic-GSQ-RCO-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by 0bserverx, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwythos-9b-claude-mythos-5-1m-abliterated-gguf", "name": "Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-25", "lastVerified": "2026-06-25", "ugiScore": 65.2, "complianceRate": 98.6, "mmluPro": 57.8, "eqBench": 76.9, "ifeval": 75.5, "hfDownloads30d": 60019, "hfLikes": 225, "ollamaCommand": "huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Open Transformer (9B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-6-27b-abliterated-mtp-gguf", "name": "Huihui-Qwen3.6-27B-abliterated-MTP-GGUF", "hfRepo": "huihui-ai/Huihui-Qwen3.6-27B-abliterated-MTP-GGUF", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-05-19", "lastVerified": "2026-05-19", "ugiScore": 73.2, "complianceRate": 97.6, "mmluPro": 68.6, "eqBench": 82.1, "ifeval": 84.1, "hfDownloads30d": 60934, "hfLikes": 100, "ollamaCommand": "huihui-ai/Huihui-Qwen3.6-27B-abliterated-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.6 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "os-software-ternary-bonsai-2-27b-uncensored-heretic-v2-gguf", "name": "Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-GGUF", "hfRepo": "OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-GGUF", "creator": "OS-Software", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-03", "lastVerified": "2026-11-01", "ugiScore": 69.8, "complianceRate": 97.6, "mmluPro": 65.5, "eqBench": 82.6, "ifeval": 80.2, "hfDownloads30d": 61177, "hfLikes": 34, "ollamaCommand": "OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by OS-Software, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-3-llama-3-1-8b-gguf", "name": "Hermes-3-Llama-3.1-8B-GGUF", "hfRepo": "NousResearch/Hermes-3-Llama-3.1-8B-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-05", "lastVerified": "2024-08-23", "ugiScore": 67.3, "complianceRate": 99.5, "mmluPro": 56.1, "eqBench": 79.7, "ifeval": 77.9, "hfDownloads30d": 58053, "hfLikes": 223, "ollamaCommand": "NousResearch/Hermes-3-Llama-3.1-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "hauhaucs-qwen3-5-2b-uncensored-hauhaucs-aggressive", "name": "Qwen3.5-2B-Uncensored-HauhauCS-Aggressive", "hfRepo": "HauhauCS/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive", "creator": "HauhauCS", "baseArchitecture": "Qwen 3.5 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-03-05", "lastVerified": "2026-11-01", "ugiScore": 62.6, "complianceRate": 99.4, "mmluPro": 50.4, "eqBench": 70, "ifeval": 72.2, "hfDownloads30d": 57614, "hfLikes": 174, "ollamaCommand": "HauhauCS/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.5 (2B) by HauhauCS, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.4 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 2B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "blackfrost-ai-qwen3-8-27b-abliterated-gguf", "name": "Qwen3.8-27B-ABLITERATED-GGUF", "hfRepo": "Blackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF", "creator": "Blackfrost-AI", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-14", "lastVerified": "2026-11-01", "ugiScore": 71.4, "complianceRate": 98.8, "mmluPro": 67.4, "eqBench": 81.3, "ifeval": 82.3, "hfDownloads30d": 55684, "hfLikes": 266, "ollamaCommand": "Blackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by Blackfrost-AI, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "istrosec-thinkingcap-qwen3-8-27b-abliterated-nvfp4a16", "name": "ThinkingCap-Qwen3.8-27B-abliterated-NVFP4A16", "hfRepo": "IstroSec/ThinkingCap-Qwen3.8-27B-abliterated-NVFP4A16", "creator": "IstroSec", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-25", "lastVerified": "2026-11-01", "ugiScore": 72, "complianceRate": 97.6, "mmluPro": 67.1, "eqBench": 81.7, "ifeval": 82.9, "hfDownloads30d": 55060, "hfLikes": 1, "ollamaCommand": "IstroSec/ThinkingCap-Qwen3.8-27B-abliterated-NVFP4A16", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by IstroSec, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-6-35b-a3b-claude-4-7-opus-abliterated-mtp-gguf", "name": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MTP-GGUF", "hfRepo": "huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MTP-GGUF", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-05-19", "lastVerified": "2026-05-19", "ugiScore": 73.5, "complianceRate": 98.5, "mmluPro": 69.4, "eqBench": 82.5, "ifeval": 84.4, "hfDownloads30d": 50531, "hfLikes": 124, "ollamaCommand": "huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated-MTP-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sc117-qwen3-6-35b-a3b-uncensored-heretic-native-mtp-preserved-apex-gguf", "name": "Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-GGUF", "hfRepo": "SC117/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-GGUF", "creator": "SC117", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-05-26", "lastVerified": "2026-11-01", "ugiScore": 73.9, "complianceRate": 98.9, "mmluPro": 69.1, "eqBench": 82.5, "ifeval": 84.8, "hfDownloads30d": 49961, "hfLikes": 157, "ollamaCommand": "SC117/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-APEX-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by SC117, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin3-0-llama3-1-8b-gguf", "name": "Dolphin3.0-Llama3.1-8B-GGUF", "hfRepo": "dphn/Dolphin3.0-Llama3.1-8B-GGUF", "creator": "dphn", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-01-02", "lastVerified": "2026-11-01", "ugiScore": 65.2, "complianceRate": 98.9, "mmluPro": 55.4, "eqBench": 75.8, "ifeval": 74.5, "hfDownloads30d": 48428, "hfLikes": 249, "ollamaCommand": "dphn/Dolphin3.0-Llama3.1-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "davidau-llama-3-2-8x3b-moe-dark-champion-instruct-uncensored-abliterated-18-4b-gguf", "name": "Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF", "hfRepo": "DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF", "creator": "DavidAU", "baseArchitecture": "Llama 3.2 (18.4B)", "paramsBillions": 18.4, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-12-10", "lastVerified": "2026-11-01", "ugiScore": 68.1, "complianceRate": 97.6, "mmluPro": 62.2, "eqBench": 79.3, "ifeval": 78.4, "hfDownloads30d": 39077, "hfLikes": 670, "ollamaCommand": "DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF", "vramQ4GB": 12, "vramQ8GB": 21, "vramFP16GB": 40, "summary": "Orthogonal weight-projected Llama 3.2 (18.4B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (18.4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 18.4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "paperscarecrow-gemma-4-31b-it-abliterated", "name": "Gemma-4-31B-it-abliterated", "hfRepo": "paperscarecrow/Gemma-4-31B-it-abliterated", "creator": "paperscarecrow", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-02", "lastVerified": "2026-11-01", "ugiScore": 72.1, "complianceRate": 98, "mmluPro": 66.5, "eqBench": 82.7, "ifeval": 82.5, "hfDownloads30d": 46939, "hfLikes": 126, "ollamaCommand": "paperscarecrow/Gemma-4-31B-it-abliterated", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by paperscarecrow, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-12b-it-qat-q4-0-unquantized-abliterated-gguf", "name": "Huihui-gemma-4-12B-it-qat-q4_0-unquantized-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-gemma-4-12B-it-qat-q4_0-unquantized-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-08", "lastVerified": "2026-06-08", "ugiScore": 67.6, "complianceRate": 98.5, "mmluPro": 58.8, "eqBench": 79.2, "ifeval": 77.9, "hfDownloads30d": 44880, "hfLikes": 74, "ollamaCommand": "huihui-ai/Huihui-gemma-4-12B-it-qat-q4_0-unquantized-abliterated-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 2 (12B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-8-27b-abliterated", "name": "Huihui-Qwen3.8-27B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.8-27B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-16", "lastVerified": "2026-08-24", "ugiScore": 73.4, "complianceRate": 99, "mmluPro": 67.3, "eqBench": 81.2, "ifeval": 84.3, "hfDownloads30d": 35417, "hfLikes": 676, "ollamaCommand": "huihui-ai/Huihui-Qwen3.8-27B-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "roadtonowhere-qwen3-32b-abliterated-q4-k-m-gguf", "name": "Qwen3-32B-abliterated-Q4_K_M-GGUF", "hfRepo": "RoadToNowhere/Qwen3-32B-abliterated-Q4_K_M-GGUF", "creator": "RoadToNowhere", "baseArchitecture": "Qwen 3 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-07", "lastVerified": "2026-11-01", "ugiScore": 73.7, "complianceRate": 97.9, "mmluPro": 69, "eqBench": 82.6, "ifeval": 84.6, "hfDownloads30d": 44962, "hfLikes": 3, "ollamaCommand": "RoadToNowhere/Qwen3-32B-abliterated-Q4_K_M-GGUF", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 3 (32B) checkpoint by RoadToNowhere, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "zetasepic-qwen2-5-72b-instruct-abliterated-gguf", "name": "Qwen2.5-72B-Instruct-abliterated-GGUF", "hfRepo": "zetasepic/Qwen2.5-72B-Instruct-abliterated-GGUF", "creator": "zetasepic", "baseArchitecture": "Qwen 2.5 (72B)", "paramsBillions": 72, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-10-01", "lastVerified": "2026-11-01", "ugiScore": 76, "complianceRate": 97.7, "mmluPro": 73.8, "eqBench": 85.1, "ifeval": 87, "hfDownloads30d": 43446, "hfLikes": 7, "ollamaCommand": "zetasepic/Qwen2.5-72B-Instruct-abliterated-GGUF", "vramQ4GB": 42, "vramQ8GB": 79, "vramFP16GB": 155, "summary": "Orthogonal weight-projected Qwen 2.5 (72B) checkpoint by zetasepic, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~42 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (72B) instruction and coding schemas" ], "tradeoffs": [ "Requires 42 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-8-27b-coder390-efficientthink-abliterated-gguf", "name": "Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-14", "lastVerified": "2026-10-14", "ugiScore": 73, "complianceRate": 99.2, "mmluPro": 69.5, "eqBench": 78.9, "ifeval": 86.1, "hfDownloads30d": 42515, "hfLikes": 28, "ollamaCommand": "huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-rocinante-x-12b-v1-gguf", "name": "Rocinante-X-12B-v1-GGUF", "hfRepo": "TheDrummer/Rocinante-X-12B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-01-19", "lastVerified": "2026-01-25", "ugiScore": 67.6, "complianceRate": 98.9, "mmluPro": 57.2, "eqBench": 81.6, "ifeval": 77.9, "hfDownloads30d": 42157, "hfLikes": 42, "ollamaCommand": "TheDrummer/Rocinante-X-12B-v1-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "groxaxo-huihui-gemma-4-26b-a4b-it-abliterated-gguf", "name": "Huihui-gemma-4-26B-A4B-it-abliterated-GGUF", "hfRepo": "groxaxo/Huihui-gemma-4-26B-A4B-it-abliterated-GGUF", "creator": "groxaxo", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-11", "lastVerified": "2026-11-01", "ugiScore": 70.1, "complianceRate": 99.2, "mmluPro": 65.4, "eqBench": 82.4, "ifeval": 80.4, "hfDownloads30d": 41971, "hfLikes": 51, "ollamaCommand": "groxaxo/Huihui-gemma-4-26B-A4B-it-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by groxaxo, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "pottokao-minimax-h3-textencoder-qwen3vl-32b-abliterated-gguf", "name": "MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-GGUF", "hfRepo": "pottokao/MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-GGUF", "creator": "pottokao", "baseArchitecture": "Qwen 3 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-05", "lastVerified": "2026-11-01", "ugiScore": 72.3, "complianceRate": 97.7, "mmluPro": 72.3, "eqBench": 79.8, "ifeval": 85.4, "hfDownloads30d": 42276, "hfLikes": 14, "ollamaCommand": "pottokao/MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-GGUF", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 3 (32B) checkpoint by pottokao, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-27b-abliterated", "name": "Huihui-Qwen3.5-27B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-27B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-02-27", "lastVerified": "2026-03-02", "ugiScore": 73.1, "complianceRate": 98.1, "mmluPro": 68.4, "eqBench": 82.7, "ifeval": 84, "hfDownloads30d": 39353, "hfLikes": 163, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-27B-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.5 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-meta-llama-3-1-8b-instruct-abliterated-gguf", "name": "Meta-Llama-3.1-8B-Instruct-abliterated-GGUF", "hfRepo": "mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated-GGUF", "creator": "mlabonne", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-07-24", "lastVerified": "2024-08-03", "ugiScore": 68, "complianceRate": 98.6, "mmluPro": 57.3, "eqBench": 78.1, "ifeval": 79.5, "hfDownloads30d": 37793, "hfLikes": 243, "ollamaCommand": "mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3.1 (8B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.1 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "gbuzhf-ornith-1-5-35b-a3b-abliterated-mtp-ud-apex-gguf", "name": "Ornith-1.5-35B-A3B-Abliterated-MTP-UD-APEX-GGUF", "hfRepo": "gbuzhf/Ornith-1.5-35B-A3B-Abliterated-MTP-UD-APEX-GGUF", "creator": "gbuzhf", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-22", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 99, "mmluPro": 67.2, "eqBench": 82.6, "ifeval": 82.8, "hfDownloads30d": 40260, "hfLikes": 43, "ollamaCommand": "gbuzhf/Ornith-1.5-35B-A3B-Abliterated-MTP-UD-APEX-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Open Transformer (35B) checkpoint by gbuzhf, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davidau-qwen3-8-27b-twin-turbo-fable-cold-fusion-709-ultra-heretic-uncensored-nm-dau-neo-mtp-gguf", "name": "Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NM-DAU-NEO-MTP-GGUF", "hfRepo": "DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NM-DAU-NEO-MTP-GGUF", "creator": "DavidAU", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-17", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.7, "mmluPro": 68, "eqBench": 81.7, "ifeval": 83.4, "hfDownloads30d": 39339, "hfLikes": 92, "ollamaCommand": "DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-ULTRA-HERETIC-Uncensored-NM-DAU-NEO-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "emiltsoi-ternary-bonsai-2-27b-uncensored-heretic-ninfer", "name": "Ternary-Bonsai-2-27B-Uncensored-Heretic-NInfer", "hfRepo": "emiltsoi/Ternary-Bonsai-2-27B-Uncensored-Heretic-NInfer", "creator": "emiltsoi", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-25", "lastVerified": "2026-11-01", "ugiScore": 69.8, "complianceRate": 98.2, "mmluPro": 64.3, "eqBench": 81.3, "ifeval": 80.2, "hfDownloads30d": 40495, "hfLikes": 12, "ollamaCommand": "emiltsoi/Ternary-Bonsai-2-27B-Uncensored-Heretic-NInfer", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by emiltsoi, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "boldingbuilds-ternary-bonsai-2-27b-abliterated-v2-pq2-0-mtp-gguf", "name": "Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-GGUF", "hfRepo": "BoldingBuilds/Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-GGUF", "creator": "BoldingBuilds", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-26", "lastVerified": "2026-11-01", "ugiScore": 72.1, "complianceRate": 98.7, "mmluPro": 64.6, "eqBench": 82, "ifeval": 82.5, "hfDownloads30d": 39698, "hfLikes": 38, "ollamaCommand": "BoldingBuilds/Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by BoldingBuilds, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "barding-defense-qwen3-8-27b-huihui-abliterated-nvfp4-ninfer", "name": "Qwen3.8-27B-huihui-abliterated-NVFP4-NInfer", "hfRepo": "Barding-Defense/Qwen3.8-27B-huihui-abliterated-NVFP4-NInfer", "creator": "Barding-Defense", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-03", "lastVerified": "2026-11-01", "ugiScore": 72.9, "complianceRate": 97.9, "mmluPro": 67.6, "eqBench": 81.3, "ifeval": 83.8, "hfDownloads30d": 38371, "hfLikes": 12, "ollamaCommand": "Barding-Defense/Qwen3.8-27B-huihui-abliterated-NVFP4-NInfer", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by Barding-Defense, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "abliterant-qwen3-8-27b-rana-abliterated-gguf", "name": "Qwen3.8-27B-RANA-abliterated-GGUF", "hfRepo": "abliterant/Qwen3.8-27B-RANA-abliterated-GGUF", "creator": "abliterant", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-26", "lastVerified": "2026-11-01", "ugiScore": 73.3, "complianceRate": 98.3, "mmluPro": 67.3, "eqBench": 82, "ifeval": 84.2, "hfDownloads30d": 37761, "hfLikes": 2, "ollamaCommand": "abliterant/Qwen3.8-27B-RANA-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by abliterant, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "the-freezer-muse-glimmer-30b-abliterated-gguf", "name": "Muse-Glimmer-30B-Abliterated-GGUF", "hfRepo": "The-freezer/Muse-Glimmer-30B-Abliterated-GGUF", "creator": "The-freezer", "baseArchitecture": "Open Transformer (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-10", "lastVerified": "2026-11-01", "ugiScore": 72.2, "complianceRate": 98.3, "mmluPro": 65.6, "eqBench": 83.2, "ifeval": 82.6, "hfDownloads30d": 36226, "hfLikes": 56, "ollamaCommand": "The-freezer/Muse-Glimmer-30B-Abliterated-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Open Transformer (30B) checkpoint by The-freezer, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nif0-qwen3-vl-32b-instruct-ultra-uncensored-heretic-h3-gguf", "name": "Qwen3-VL-32B-Instruct-ultra-uncensored-heretic-H3-GGUF", "hfRepo": "nif0/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic-H3-GGUF", "creator": "nif0", "baseArchitecture": "Qwen 3 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-04", "lastVerified": "2026-11-01", "ugiScore": 73, "complianceRate": 97.8, "mmluPro": 69.5, "eqBench": 81.9, "ifeval": 83.9, "hfDownloads30d": 36253, "hfLikes": 44, "ollamaCommand": "nif0/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic-H3-GGUF", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 3 (32B) checkpoint by nif0, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "bennydaball-qwen3-4b-z-image-turbo-abliteratedv1", "name": "Qwen3-4b-Z-Image-Turbo-AbliteratedV1", "hfRepo": "BennyDaBall/Qwen3-4b-Z-Image-Turbo-AbliteratedV1", "creator": "BennyDaBall", "baseArchitecture": "Qwen 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-01-28", "lastVerified": "2026-11-01", "ugiScore": 63.7, "complianceRate": 99.2, "mmluPro": 54.2, "eqBench": 73.6, "ifeval": 74.4, "hfDownloads30d": 34050, "hfLikes": 130, "ollamaCommand": "BennyDaBall/Qwen3-4b-Z-Image-Turbo-AbliteratedV1", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3 (4B) checkpoint by BennyDaBall, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "noobito45-qwen3-8-9b-heretic-uncensored-nvfp4-gguf", "name": "Qwen3.8-9B-heretic-uncensored-NVFP4-GGUF", "hfRepo": "Noobito45/Qwen3.8-9B-heretic-uncensored-NVFP4-GGUF", "creator": "Noobito45", "baseArchitecture": "Qwen 3.8 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-18", "lastVerified": "2026-11-01", "ugiScore": 68.6, "complianceRate": 98, "mmluPro": 61.1, "eqBench": 77.6, "ifeval": 79.4, "hfDownloads30d": 35143, "hfLikes": 45, "ollamaCommand": "Noobito45/Qwen3.8-9B-heretic-uncensored-NVFP4-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.8 (9B) checkpoint by Noobito45, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "bucoid-qwen3-8-27b-heretic-ara-16gb-vram-iq4-xs-mtp-gguf", "name": "Qwen3.8-27B-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF", "hfRepo": "Bucoid/Qwen3.8-27B-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF", "creator": "Bucoid", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-18", "lastVerified": "2026-11-01", "ugiScore": 73.3, "complianceRate": 97.7, "mmluPro": 68.4, "eqBench": 82.6, "ifeval": 84.2, "hfDownloads30d": 32305, "hfLikes": 94, "ollamaCommand": "Bucoid/Qwen3.8-27B-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by Bucoid, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "abiray-qwen3-5-9b-abliterated-gguf", "name": "Qwen3.5-9B-abliterated-GGUF", "hfRepo": "Abiray/Qwen3.5-9B-abliterated-GGUF", "creator": "Abiray", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-06", "lastVerified": "2026-11-01", "ugiScore": 67.3, "complianceRate": 97.9, "mmluPro": 60.6, "eqBench": 77.3, "ifeval": 78.1, "hfDownloads30d": 32651, "hfLikes": 19, "ollamaCommand": "Abiray/Qwen3.5-9B-abliterated-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by Abiray, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thaurock-qwen3-coder-30b-a3b-instruct-abliterated-gguf", "name": "Qwen3-Coder-30B-A3B-Instruct-abliterated-GGUF", "hfRepo": "Thaurock/Qwen3-Coder-30B-A3B-Instruct-abliterated-GGUF", "creator": "Thaurock", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-23", "lastVerified": "2026-11-01", "ugiScore": 72.6, "complianceRate": 97.7, "mmluPro": 71.4, "eqBench": 80.6, "ifeval": 85.7, "hfDownloads30d": 31804, "hfLikes": 1, "ollamaCommand": "Thaurock/Qwen3-Coder-30B-A3B-Instruct-abliterated-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by Thaurock, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-26b-a4b-it-qat-q4-0-unquantized-abliterated-gguf", "name": "Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-11", "lastVerified": "2026-06-11", "ugiScore": 71.9, "complianceRate": 98.6, "mmluPro": 63.5, "eqBench": 81.8, "ifeval": 82.2, "hfDownloads30d": 31159, "hfLikes": 24, "ollamaCommand": "huihui-ai/Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-4-3-36b-gguf", "name": "Hermes-4.3-36B-GGUF", "hfRepo": "NousResearch/Hermes-4.3-36B-GGUF", "creator": "NousResearch", "baseArchitecture": "Open Transformer (36B)", "paramsBillions": 36, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-11-25", "lastVerified": "2025-12-03", "ugiScore": 73.4, "complianceRate": 98.7, "mmluPro": 65.8, "eqBench": 82.1, "ifeval": 83, "hfDownloads30d": 28799, "hfLikes": 97, "ollamaCommand": "NousResearch/Hermes-4.3-36B-GGUF", "vramQ4GB": 21, "vramQ8GB": 40, "vramFP16GB": 78, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (36B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 36B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-31b-it-qat-q4-0-unquantized-abliterated-gguf", "name": "Huihui-gemma-4-31B-it-qat-q4_0-unquantized-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-gemma-4-31B-it-qat-q4_0-unquantized-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-11", "lastVerified": "2026-06-11", "ugiScore": 70.7, "complianceRate": 97.8, "mmluPro": 64.9, "eqBench": 82.3, "ifeval": 81.1, "hfDownloads30d": 29498, "hfLikes": 23, "ollamaCommand": "huihui-ai/Huihui-gemma-4-31B-it-qat-q4_0-unquantized-abliterated-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sc117-gemma-4-26b-a4b-it-qat-heretic-gguf", "name": "gemma-4-26B-A4B-it-qat-heretic-GGUF", "hfRepo": "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF", "creator": "SC117", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-08", "lastVerified": "2026-11-01", "ugiScore": 71.5, "complianceRate": 97.6, "mmluPro": 64.1, "eqBench": 82.2, "ifeval": 81.8, "hfDownloads30d": 28934, "hfLikes": 40, "ollamaCommand": "SC117/gemma-4-26B-A4B-it-qat-heretic-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by SC117, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-vl-8b-instruct-abliterated", "name": "Huihui-Qwen3-VL-8B-Instruct-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-10-16", "lastVerified": "2026-11-01", "ugiScore": 66.8, "complianceRate": 99.1, "mmluPro": 59.7, "eqBench": 76.8, "ifeval": 77.5, "hfDownloads30d": 25174, "hfLikes": 225, "ollamaCommand": "huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Qwen 3 (8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sc117-ornith-1-0-35b-heretic-mtp-apex-gguf", "name": "Ornith-1.0-35B-Heretic-MTP-APEX-GGUF", "hfRepo": "SC117/Ornith-1.0-35B-Heretic-MTP-APEX-GGUF", "creator": "SC117", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-12", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 99, "mmluPro": 66.4, "eqBench": 83.5, "ifeval": 82.8, "hfDownloads30d": 25527, "hfLikes": 30, "ollamaCommand": "SC117/Ornith-1.0-35B-Heretic-MTP-APEX-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Open Transformer (35B) checkpoint by SC117, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwable-3-6-27b-abliterated-mtp-gguf", "name": "Huihui-Qwable-3.6-27b-abliterated-MTP-GGUF", "hfRepo": "huihui-ai/Huihui-Qwable-3.6-27b-abliterated-MTP-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-21", "lastVerified": "2026-06-21", "ugiScore": 72, "complianceRate": 98.6, "mmluPro": 64.8, "eqBench": 82.7, "ifeval": 82.4, "hfDownloads30d": 25402, "hfLikes": 34, "ollamaCommand": "huihui-ai/Huihui-Qwable-3.6-27b-abliterated-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-8b-abliterated-v2", "name": "Huihui-Qwen3-8B-abliterated-v2", "hfRepo": "huihui-ai/Huihui-Qwen3-8B-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-18", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 97.6, "mmluPro": 60.3, "eqBench": 76.7, "ifeval": 77.8, "hfDownloads30d": 24772, "hfLikes": 72, "ollamaCommand": "huihui-ai/Huihui-Qwen3-8B-abliterated-v2", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Qwen 3 (8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "ttimofeyka-mistralrp-noromaid-nsfw-mistral-7b-gguf", "name": "MistralRP-Noromaid-NSFW-Mistral-7B-GGUF", "hfRepo": "Ttimofeyka/MistralRP-Noromaid-NSFW-Mistral-7B-GGUF", "creator": "Ttimofeyka", "baseArchitecture": "Mistral (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-02-07", "lastVerified": "2026-11-01", "ugiScore": 64.7, "complianceRate": 98.2, "mmluPro": 54.9, "eqBench": 75.8, "ifeval": 74, "hfDownloads30d": 23434, "hfLikes": 147, "ollamaCommand": "Ttimofeyka/MistralRP-Noromaid-NSFW-Mistral-7B-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Mistral (7B) by Ttimofeyka, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwable-3-6-27b-abliterated-gguf", "name": "Huihui-Qwable-3.6-27b-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-Qwable-3.6-27b-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-19", "lastVerified": "2026-06-21", "ugiScore": 70.2, "complianceRate": 98.6, "mmluPro": 64.1, "eqBench": 81.9, "ifeval": 80.6, "hfDownloads30d": 24776, "hfLikes": 24, "ollamaCommand": "huihui-ai/Huihui-Qwable-3.6-27b-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Open Transformer (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "wangzhang-qwen3-6-35b-a3b-abliterated-gguf", "name": "Qwen3.6-35B-A3B-abliterated-GGUF", "hfRepo": "wangzhang/Qwen3.6-35B-A3B-abliterated-GGUF", "creator": "wangzhang", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-17", "lastVerified": "2026-11-01", "ugiScore": 73.1, "complianceRate": 98.1, "mmluPro": 70.5, "eqBench": 83.2, "ifeval": 84, "hfDownloads30d": 24815, "hfLikes": 5, "ollamaCommand": "wangzhang/Qwen3.6-35B-A3B-abliterated-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by wangzhang, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "andycurrent-gemma-3-1b-it-glm-4-7-flash-heretic-uncensored-thinking-gguf", "name": "Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "hfRepo": "Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "creator": "Andycurrent", "baseArchitecture": "Gemma 3 (1B)", "paramsBillions": 1, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-02-18", "lastVerified": "2026-11-01", "ugiScore": 58.7, "complianceRate": 97.9, "mmluPro": 45.8, "eqBench": 73, "ifeval": 68.8, "hfDownloads30d": 21563, "hfLikes": 167, "ollamaCommand": "Andycurrent/Gemma-3-1B-it-GLM-4.7-Flash-Heretic-Uncensored-Thinking_GGUF", "vramQ4GB": 1.7, "vramQ8GB": 2.3, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Gemma 3 (1B) checkpoint by Andycurrent, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.7 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (1B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thunder13240-ornith-1-5-9b-heretic-gguf", "name": "Ornith-1.5-9B-heretic-GGUF", "hfRepo": "Thunder13240/Ornith-1.5-9B-heretic-GGUF", "creator": "Thunder13240", "baseArchitecture": "Open Transformer (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-20", "lastVerified": "2026-11-01", "ugiScore": 65.4, "complianceRate": 98.2, "mmluPro": 57.1, "eqBench": 76.4, "ifeval": 75.7, "hfDownloads30d": 23575, "hfLikes": 27, "ollamaCommand": "Thunder13240/Ornith-1.5-9B-heretic-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Open Transformer (9B) checkpoint by Thunder13240, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-mixtral-8x7b-dpo", "name": "Nous-Hermes-2-Mixtral-8x7B-DPO", "hfRepo": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO", "creator": "NousResearch", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-11", "lastVerified": "2024-04-30", "ugiScore": 75.3, "complianceRate": 98.8, "mmluPro": 66.5, "eqBench": 84.8, "ifeval": 84.9, "hfDownloads30d": 16825, "hfLikes": 460, "ollamaCommand": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gpt-oss-20b-mxfp4-abliterated-v2", "name": "Huihui-gpt-oss-20b-mxfp4-abliterated-v2", "hfRepo": "huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (20B)", "paramsBillions": 20, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-09-27", "lastVerified": "2026-11-01", "ugiScore": 70.4, "complianceRate": 98.3, "mmluPro": 62.2, "eqBench": 80.3, "ifeval": 80.7, "hfDownloads30d": 23399, "hfLikes": 20, "ollamaCommand": "huihui-ai/Huihui-gpt-oss-20b-mxfp4-abliterated-v2", "vramQ4GB": 12, "vramQ8GB": 23, "vramFP16GB": 44, "summary": "Orthogonal weight-projected Open Transformer (20B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (20B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 20B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "0bserverx-muse-glimmer-30b-heretic-uncensored-gguf", "name": "Muse-Glimmer-30B-Heretic-Uncensored-GGUF", "hfRepo": "0bserverx/Muse-Glimmer-30B-Heretic-Uncensored-GGUF", "creator": "0bserverx", "baseArchitecture": "Open Transformer (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-10", "lastVerified": "2026-11-01", "ugiScore": 71, "complianceRate": 98.9, "mmluPro": 65.1, "eqBench": 81.3, "ifeval": 81.4, "hfDownloads30d": 22264, "hfLikes": 30, "ollamaCommand": "0bserverx/Muse-Glimmer-30B-Heretic-Uncensored-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Open Transformer (30B) checkpoint by 0bserverx, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-coder-30b-a3b-instruct-abliterated", "name": "Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-08-02", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.8, "mmluPro": 71.4, "eqBench": 79.2, "ifeval": 85.6, "hfDownloads30d": 21193, "hfLikes": 72, "ollamaCommand": "huihui-ai/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mudler-gemma-4-26b-a4b-it-heretic-apex-gguf", "name": "gemma-4-26B-A4B-it-heretic-APEX-GGUF", "hfRepo": "mudler/gemma-4-26B-A4B-it-heretic-APEX-GGUF", "creator": "mudler", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-05", "lastVerified": "2026-11-01", "ugiScore": 71.9, "complianceRate": 98, "mmluPro": 64.7, "eqBench": 81.7, "ifeval": 82.2, "hfDownloads30d": 21402, "hfLikes": 56, "ollamaCommand": "mudler/gemma-4-26B-A4B-it-heretic-APEX-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by mudler, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "qwen-2.5-coder-32b-abliterated", "name": "Qwen2.5-Coder-32B-Instruct-abliterated", "hfRepo": "huihui-ai/Qwen2.5-Coder-32B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 Coder", "paramsBillions": 32.5, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-11-14", "lastVerified": "2025-05-10", "ugiScore": 73.9, "complianceRate": 98.5, "mmluPro": 73.1, "eqBench": 78.4, "ifeval": 87.3, "hfDownloads30d": 15940, "hfLikes": 415, "openRouterSlug": "qwen/qwen-2.5-coder-32b-instruct", "ollamaCommand": "huihui-ai/Qwen2.5-Coder-32B-Instruct-abliterated", "vramQ4GB": 20, "vramQ8GB": 35, "vramFP16GB": 65, "summary": "The gold standard for local offensive security, exploit analysis, reverse engineering, and unrestricted software architecture on a single 24GB GPU.", "strengths": [ "Top open-weight 32B coding & reverse-engineering accuracy", "No refusals on penetration testing or CVE research" ], "tradeoffs": [ "Less suited for literary fiction than EVA or Magnum" ] }, { "id": "hotdogs-qwen3-8-27b-abliterated-cyber-preview-mtp-gguf", "name": "Qwen3.8-27B-abliterated-cyber-preview-MTP-GGUF", "hfRepo": "hotdogs/Qwen3.8-27B-abliterated-cyber-preview-MTP-GGUF", "creator": "hotdogs", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-19", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.7, "mmluPro": 67.1, "eqBench": 82.7, "ifeval": 83.4, "hfDownloads30d": 21048, "hfLikes": 8, "ollamaCommand": "hotdogs/Qwen3.8-27B-abliterated-cyber-preview-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by hotdogs, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davetha-qwen3-8-27b-abliterated-w8a16", "name": "Qwen3.8-27B-abliterated-W8A16", "hfRepo": "davetha/Qwen3.8-27B-abliterated-W8A16", "creator": "davetha", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-14", "lastVerified": "2026-11-01", "ugiScore": 71.7, "complianceRate": 97.9, "mmluPro": 68.7, "eqBench": 81.1, "ifeval": 82.6, "hfDownloads30d": 20896, "hfLikes": 1, "ollamaCommand": "davetha/Qwen3.8-27B-abliterated-W8A16", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by davetha, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-neuraldaredevil-8b-abliterated", "name": "NeuralDaredevil-8B-abliterated", "hfRepo": "mlabonne/NeuralDaredevil-8B-abliterated", "creator": "mlabonne", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-27", "lastVerified": "2026-01-23", "ugiScore": 65.5, "complianceRate": 98.2, "mmluPro": 57.3, "eqBench": 76.6, "ifeval": 75.7, "hfDownloads30d": 16099, "hfLikes": 299, "ollamaCommand": "mlabonne/NeuralDaredevil-8B-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Open Transformer (8B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "douganger-dolphin-mistral-24b-venice-edition-1-2-gguf", "name": "Dolphin-Mistral-24B-Venice-Edition-1.2-GGUF", "hfRepo": "douganger/Dolphin-Mistral-24B-Venice-Edition-1.2-GGUF", "creator": "douganger", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-31", "lastVerified": "2026-11-01", "ugiScore": 71, "complianceRate": 98.1, "mmluPro": 64.1, "eqBench": 81.6, "ifeval": 80.5, "hfDownloads30d": 20563, "hfLikes": 1, "ollamaCommand": "douganger/Dolphin-Mistral-24B-Venice-Edition-1.2-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored DPO/SFT fine-tune of Mistral Small 3 (24B) by douganger, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 24B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "criscar22-qwen2-5-coder-7b-instruct-abliterated", "name": "Qwen2.5-Coder-7B-Instruct-abliterated", "hfRepo": "criscar22/Qwen2.5-Coder-7B-Instruct-abliterated", "creator": "criscar22", "baseArchitecture": "Qwen 2.5 Coder (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-27", "lastVerified": "2026-11-01", "ugiScore": 67.5, "complianceRate": 99.2, "mmluPro": 62.4, "eqBench": 74.4, "ifeval": 80.4, "hfDownloads30d": 20341, "hfLikes": 1, "ollamaCommand": "criscar22/Qwen2.5-Coder-7B-Instruct-abliterated", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (7B) checkpoint by criscar22, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "llama-3.3-70b-instruct-abliterated", "name": "Llama-3.3-70B-Instruct-abliterated", "hfRepo": "huihui-ai/Llama-3.3-70B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Llama 3.3 (Dense)", "paramsBillions": 70.6, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-12-08", "lastVerified": "2026-11-01", "ugiScore": 75.8, "complianceRate": 97.6, "mmluPro": 70.9, "eqBench": 85.4, "ifeval": 88.1, "hfDownloads30d": 18859, "hfLikes": 94, "openRouterSlug": "meta-llama/llama-3.3-70b-instruct", "ollamaCommand": "huihui-ai/Llama-3.3-70B-Instruct-abliterated", "vramQ4GB": 41, "vramQ8GB": 74, "vramFP16GB": 142, "summary": "Layer-targeted weight orthogonalization of Meta Llama 3.3 70B preserving its 88.1 IFEval instruction following score while stripping preachy refusals.", "strengths": [ "Highest IFEval instruction adherence in 70B class", "Strong English prose & dialogue" ], "tradeoffs": [ "Requires 40GB+ VRAM for Q4_K_M quantization" ], "isLiveSynced": true }, { "id": "igorls-gemma-4-12b-it-heretic-gguf", "name": "gemma-4-12B-it-heretic-GGUF", "hfRepo": "igorls/gemma-4-12B-it-heretic-GGUF", "creator": "igorls", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-04", "lastVerified": "2026-11-01", "ugiScore": 67.2, "complianceRate": 99.3, "mmluPro": 60.2, "eqBench": 78.3, "ifeval": 77.5, "hfDownloads30d": 18688, "hfLikes": 67, "ollamaCommand": "igorls/gemma-4-12B-it-heretic-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 2 (12B) checkpoint by igorls, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-mistral-24b-venice-edition", "name": "Dolphin-Mistral-24B-Venice-Edition", "hfRepo": "dphn/Dolphin-Mistral-24B-Venice-Edition", "creator": "dphn", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-06-12", "lastVerified": "2026-11-01", "ugiScore": 70.2, "complianceRate": 98.9, "mmluPro": 63.9, "eqBench": 80.8, "ifeval": 79.7, "hfDownloads30d": 7822, "hfLikes": 769, "ollamaCommand": "dphn/Dolphin-Mistral-24B-Venice-Edition", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored DPO/SFT fine-tune of Mistral Small 3 (24B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 24B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "dolphin-3.0-r1-mistral-24b", "name": "Dolphin3.0-R1-Mistral-24B", "hfRepo": "cognitivecomputations/Dolphin3.0-R1-Mistral-24B", "creator": "cognitivecomputations", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-02-02", "lastVerified": "2025-04-25", "ugiScore": 74.1, "complianceRate": 99.1, "mmluPro": 66.4, "eqBench": 82.9, "ifeval": 84.5, "hfDownloads30d": 14200, "hfLikes": 342, "openRouterSlug": "cognitivecomputations/dolphin3.0-r1-mistral-24b", "ollamaCommand": "cognitivecomputations/Dolphin3.0-R1-Mistral-24B", "vramQ4GB": 15, "vramQ8GB": 26, "vramFP16GB": 48, "summary": "Trained from scratch on curated uncensored reasoning traces and system-prompt obedience datasets; fits comfortably on a single RTX 3090/4090.", "strengths": [ "Complete system prompt steerability", "Runs on single 16GB–24GB GPU", "Explicit reasoning chain support" ], "tradeoffs": [ "32k context ceiling compared to 128k Llama/Qwen peers" ] }, { "id": "huihui-ai-huihui-qwen3-30b-a3b-instruct-2507-abliterated", "name": "Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-08-01", "lastVerified": "2026-11-01", "ugiScore": 72.2, "complianceRate": 99.1, "mmluPro": 68.1, "eqBench": 82.2, "ifeval": 83.1, "hfDownloads30d": 18707, "hfLikes": 41, "ollamaCommand": "huihui-ai/Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gpt-oss-20b-bf16-abliterated", "name": "Huihui-gpt-oss-20b-BF16-abliterated", "hfRepo": "huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (20B)", "paramsBillions": 20, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-08-06", "lastVerified": "2026-11-01", "ugiScore": 69.9, "complianceRate": 97.8, "mmluPro": 63.1, "eqBench": 80.3, "ifeval": 80.2, "hfDownloads30d": 15589, "hfLikes": 226, "ollamaCommand": "huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated", "vramQ4GB": 12, "vramQ8GB": 23, "vramFP16GB": 44, "summary": "Orthogonal weight-projected Open Transformer (20B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (20B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 20B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "goekdeniz-guelmez-josiefied-qwen2-5-coder-14b-instruct-abliterated-v1", "name": "Josiefied-Qwen2.5-Coder-14B-Instruct-abliterated-v1", "hfRepo": "Goekdeniz-Guelmez/Josiefied-Qwen2.5-Coder-14B-Instruct-abliterated-v1", "creator": "Goekdeniz-Guelmez", "baseArchitecture": "Qwen 2.5 Coder (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-11", "lastVerified": "2026-11-01", "ugiScore": 69, "complianceRate": 97.6, "mmluPro": 65.8, "eqBench": 76.1, "ifeval": 82, "hfDownloads30d": 18501, "hfLikes": 5, "ollamaCommand": "Goekdeniz-Guelmez/Josiefied-Qwen2.5-Coder-14B-Instruct-abliterated-v1", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (14B) checkpoint by Goekdeniz-Guelmez, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-e4b-it-qat-q4-0-unquantized-abliterated-gguf", "name": "Huihui-gemma-4-E4B-it-qat-q4_0-unquantized-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-gemma-4-E4B-it-qat-q4_0-unquantized-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-10", "lastVerified": "2026-06-11", "ugiScore": 61.4, "complianceRate": 98.5, "mmluPro": 51.4, "eqBench": 72.8, "ifeval": 71.6, "hfDownloads30d": 17865, "hfLikes": 30, "ollamaCommand": "huihui-ai/Huihui-gemma-4-E4B-it-qat-q4_0-unquantized-abliterated-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Gemma 2 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-qwen2-5-vl-3b-instruct-abliterated", "name": "Qwen2.5-VL-3B-Instruct-abliterated", "hfRepo": "huihui-ai/Qwen2.5-VL-3B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-02-16", "lastVerified": "2026-11-01", "ugiScore": 63.7, "complianceRate": 99.3, "mmluPro": 53.7, "eqBench": 72.9, "ifeval": 74.3, "hfDownloads30d": 17322, "hfLikes": 35, "ollamaCommand": "huihui-ai/Qwen2.5-VL-3B-Instruct-abliterated", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Orthogonal weight-projected Qwen 2.5 (3B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.9 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (3B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 3B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-72b-qwen2-5-kunou-v1", "name": "72B-Qwen2.5-Kunou-v1", "hfRepo": "Sao10K/72B-Qwen2.5-Kunou-v1", "creator": "Sao10K", "baseArchitecture": "Qwen 2.5 (72B)", "paramsBillions": 72, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-12-06", "lastVerified": "2024-12-15", "ugiScore": 77.9, "complianceRate": 98.8, "mmluPro": 73.4, "eqBench": 88.9, "ifeval": 88.9, "hfDownloads30d": 16922, "hfLikes": 27, "ollamaCommand": "Sao10K/72B-Qwen2.5-Kunou-v1", "vramQ4GB": 42, "vramQ8GB": 79, "vramFP16GB": 155, "summary": "Uncensored creative and roleplay specialist built on Qwen 2.5 (72B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~42 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "ge525-qwen3-8-9b-distill-uncensored-heretic-q4-k-m-gguf", "name": "Qwen3.8-9B-Distill-uncensored-heretic-Q4_K_M-GGUF", "hfRepo": "ge525/Qwen3.8-9B-Distill-uncensored-heretic-Q4_K_M-GGUF", "creator": "ge525", "baseArchitecture": "Qwen 3.8 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-20", "lastVerified": "2026-11-01", "ugiScore": 68.4, "complianceRate": 97.8, "mmluPro": 60.9, "eqBench": 77.9, "ifeval": 79.2, "hfDownloads30d": 16230, "hfLikes": 35, "ollamaCommand": "ge525/Qwen3.8-9B-Distill-uncensored-heretic-Q4_K_M-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.8 (9B) checkpoint by ge525, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-4b-abliterated", "name": "Huihui-Qwen3.5-4B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-4B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-07", "lastVerified": "2026-03-07", "ugiScore": 63.7, "complianceRate": 98.6, "mmluPro": 55.9, "eqBench": 74.6, "ifeval": 74.4, "hfDownloads30d": 15910, "hfLikes": 52, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-4B-abliterated", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3.5 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "orenguteng-llama-3-8b-lexi-uncensored-gguf", "name": "Llama-3-8B-Lexi-Uncensored-GGUF", "hfRepo": "Orenguteng/Llama-3-8B-Lexi-Uncensored-GGUF", "creator": "Orenguteng", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-04-23", "lastVerified": "2026-11-01", "ugiScore": 65.8, "complianceRate": 98.7, "mmluPro": 55.3, "eqBench": 76.3, "ifeval": 75.1, "hfDownloads30d": 11855, "hfLikes": 257, "ollamaCommand": "Orenguteng/Llama-3-8B-Lexi-Uncensored-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by Orenguteng, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "muyuxiao-qwen3-8-27b-heretic-abliterated-uncensored-gguf", "name": "Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "hfRepo": "Muyuxiao/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "creator": "Muyuxiao", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-14", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.7, "mmluPro": 67.5, "eqBench": 81.6, "ifeval": 83.4, "hfDownloads30d": 15638, "hfLikes": 1, "ollamaCommand": "Muyuxiao/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by Muyuxiao, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "gemma-2-27b-it-abliterated", "name": "gemma-2-27b-it-abliterated", "hfRepo": "mlabonne/gemma-2-27b-it-abliterated", "creator": "mlabonne", "baseArchitecture": "Gemma 2 (Dense 27B)", "paramsBillions": 27.2, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-07-18", "lastVerified": "2025-03-12", "ugiScore": 71.2, "complianceRate": 97.8, "mmluPro": 64.5, "eqBench": 85.1, "ifeval": 82.4, "hfDownloads30d": 9820, "hfLikes": 368, "openRouterSlug": "google/gemma-2-27b-it", "ollamaCommand": "mlabonne/gemma-2-27b-it-abliterated", "vramQ4GB": 17, "vramQ8GB": 29, "vramFP16GB": 55, "summary": "Reference orthogonal activation abliteration by Maxime Labonne on Gemma 2 27B, offering exceptional conversational tone in a 16GB–24GB footprint.", "strengths": [ "Punchy, natural writing style", "Fits 24GB GPU with comfortable KV cache" ], "tradeoffs": [ "8,192 token context window limit" ] }, { "id": "azminawwar-huihui-ai-huihui-qwen3-5-0-8b-abliterated-ultimate-offensive-red-team", "name": "huihui-ai_Huihui-Qwen3.5-0.8B-abliterated_Ultimate-Offensive-Red-Team", "hfRepo": "azminawwar/huihui-ai_Huihui-Qwen3.5-0.8B-abliterated_Ultimate-Offensive-Red-Team", "creator": "azminawwar", "baseArchitecture": "Qwen 3.5 (0.8B)", "paramsBillions": 0.8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-02", "lastVerified": "2026-11-01", "ugiScore": 60.2, "complianceRate": 99, "mmluPro": 48.5, "eqBench": 69.1, "ifeval": 70.8, "hfDownloads30d": 15232, "hfLikes": 5, "ollamaCommand": "azminawwar/huihui-ai_Huihui-Qwen3.5-0.8B-abliterated_Ultimate-Offensive-Red-Team", "vramQ4GB": 1.6, "vramQ8GB": 2, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Qwen 3.5 (0.8B) checkpoint by azminawwar, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.6 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (0.8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 0.8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-theta-llama-3-8b", "name": "Hermes-2-Theta-Llama-3-8B", "hfRepo": "NousResearch/Hermes-2-Theta-Llama-3-8B", "creator": "NousResearch", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-05", "lastVerified": "2024-09-08", "ugiScore": 65.5, "complianceRate": 99.2, "mmluPro": 56.1, "eqBench": 76.8, "ifeval": 74.8, "hfDownloads30d": 12135, "hfLikes": 207, "ollamaCommand": "NousResearch/Hermes-2-Theta-Llama-3-8B", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "kaushikvira-qwen3-8-27b-swift-abliterated-nvfp4full-dflash2-ninfer-v3", "name": "Qwen3.8-27B-swift-abliterated-nvfp4full-dflash2-NInfer-v3", "hfRepo": "kaushikvira/Qwen3.8-27B-swift-abliterated-nvfp4full-dflash2-NInfer-v3", "creator": "kaushikvira", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-25", "lastVerified": "2026-11-01", "ugiScore": 72.6, "complianceRate": 98.8, "mmluPro": 68.4, "eqBench": 82.6, "ifeval": 83.5, "hfDownloads30d": 14906, "hfLikes": 1, "ollamaCommand": "kaushikvira/Qwen3.8-27B-swift-abliterated-nvfp4full-dflash2-NInfer-v3", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by kaushikvira, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-27b-it-abliterated-gguf", "name": "gemma-3-27b-it-abliterated-GGUF", "hfRepo": "mlabonne/gemma-3-27b-it-abliterated-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-03-17", "lastVerified": "2025-04-01", "ugiScore": 71.1, "complianceRate": 98.9, "mmluPro": 64.2, "eqBench": 84.6, "ifeval": 81.5, "hfDownloads30d": 10864, "hfLikes": 252, "ollamaCommand": "mlabonne/gemma-3-27b-it-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Gemma 3 (27B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-anubis-mini-8b-v1-gguf", "name": "Anubis-Mini-8B-v1-GGUF", "hfRepo": "TheDrummer/Anubis-Mini-8B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-01-01", "lastVerified": "2026-03-17", "ugiScore": 66.3, "complianceRate": 98.4, "mmluPro": 54.7, "eqBench": 79.7, "ifeval": 76.5, "hfDownloads30d": 14156, "hfLikes": 21, "ollamaCommand": "TheDrummer/Anubis-Mini-8B-v1-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (8B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-ornith-1-0-9b-abliterated-mtp-gguf", "name": "Huihui-Ornith-1.0-9B-abliterated-MTP-GGUF", "hfRepo": "huihui-ai/Huihui-Ornith-1.0-9B-abliterated-MTP-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-01", "lastVerified": "2026-07-05", "ugiScore": 66.8, "complianceRate": 98.4, "mmluPro": 58.1, "eqBench": 78, "ifeval": 77.1, "hfDownloads30d": 13406, "hfLikes": 30, "ollamaCommand": "huihui-ai/Huihui-Ornith-1.0-9B-abliterated-MTP-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Open Transformer (9B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-solar-10-7b", "name": "Nous-Hermes-2-SOLAR-10.7B", "hfRepo": "NousResearch/Nous-Hermes-2-SOLAR-10.7B", "creator": "NousResearch", "baseArchitecture": "Open Transformer (10.7B)", "paramsBillions": 10.7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-01", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 99.5, "mmluPro": 58.4, "eqBench": 77.2, "ifeval": 76.5, "hfDownloads30d": 10670, "hfLikes": 211, "ollamaCommand": "NousResearch/Nous-Hermes-2-SOLAR-10.7B", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 24, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (10.7B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 10.7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-gemmasutra-mini-2b-v1-gguf", "name": "Gemmasutra-Mini-2B-v1-GGUF", "hfRepo": "TheDrummer/Gemmasutra-Mini-2B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2024-08-03", "lastVerified": "2026-11-01", "ugiScore": 58, "complianceRate": 99.2, "mmluPro": 45.2, "eqBench": 74.1, "ifeval": 68, "hfDownloads30d": 12306, "hfLikes": 90, "ollamaCommand": "TheDrummer/Gemmasutra-Mini-2B-v1-GGUF", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (2B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~2.4 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "wavecut-huihui-qwen3-8-27b-abliterated-ninfer-v3", "name": "Huihui-Qwen3.8-27B-abliterated-NInfer-v3", "hfRepo": "WaveCut/Huihui-Qwen3.8-27B-abliterated-NInfer-v3", "creator": "WaveCut", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-21", "lastVerified": "2026-11-01", "ugiScore": 72.1, "complianceRate": 98.3, "mmluPro": 67.1, "eqBench": 80.9, "ifeval": 83, "hfDownloads30d": 13510, "hfLikes": 8, "ollamaCommand": "WaveCut/Huihui-Qwen3.8-27B-abliterated-NInfer-v3", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by WaveCut, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "llama-3.1-8b-lexi-uncensored-v2", "name": "Llama-3.1-8B-Lexi-Uncensored-V2", "hfRepo": "Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2", "creator": "Orenguteng", "baseArchitecture": "Llama 3.1 (Dense 8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-21", "lastVerified": "2026-11-01", "ugiScore": 67.4, "complianceRate": 99.2, "mmluPro": 56.8, "eqBench": 79.2, "ifeval": 80.5, "hfDownloads30d": 8460, "hfLikes": 339, "ollamaCommand": "Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2", "vramQ4GB": 5.5, "vramQ8GB": 9, "vramFP16GB": 16, "summary": "Most downloaded small uncensored model on Hugging Face; runs at 90+ tok/s on consumer laptops and 8GB GPUs without refusal guardrails.", "strengths": [ "Runs on 6GB–8GB VRAM laptops & MacBooks", "99.2% compliance without coherence loss" ], "tradeoffs": [ "8B parameter ceiling limits multi-step STEM synthesis" ], "isLiveSynced": true }, { "id": "yachen4ever-qwen3-8-4b-distill-heretic-abliterated-gguf", "name": "Qwen3.8-4B-Distill-Heretic-Abliterated-GGUF", "hfRepo": "yachen4ever/Qwen3.8-4B-Distill-Heretic-Abliterated-GGUF", "creator": "yachen4ever", "baseArchitecture": "Qwen 3.8 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-23", "lastVerified": "2026-11-01", "ugiScore": 63.1, "complianceRate": 98.6, "mmluPro": 55.6, "eqBench": 73.6, "ifeval": 73.8, "hfDownloads30d": 13297, "hfLikes": 16, "ollamaCommand": "yachen4ever/Qwen3.8-4B-Distill-Heretic-Abliterated-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3.8 (4B) checkpoint by yachen4ever, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "soyaakinohara-qwen3-8-27b-abliterated-3-69bpw-12gb-mtp-gguf", "name": "qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf", "hfRepo": "soyaakinohara/qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf", "creator": "soyaakinohara", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-16", "lastVerified": "2026-11-01", "ugiScore": 73.4, "complianceRate": 98.4, "mmluPro": 68.2, "eqBench": 81.1, "ifeval": 84.3, "hfDownloads30d": 11755, "hfLikes": 114, "ollamaCommand": "soyaakinohara/qwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by soyaakinohara, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "slickpulla-qwen3-8-27b-heretic-abliterated-uncensored-gguf", "name": "Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "hfRepo": "SlickPulla/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "creator": "SlickPulla", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-04", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 98.6, "mmluPro": 68.4, "eqBench": 81.6, "ifeval": 83.3, "hfDownloads30d": 13355, "hfLikes": 1, "ollamaCommand": "SlickPulla/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by SlickPulla, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-yi-34b", "name": "Nous-Hermes-2-Yi-34B", "hfRepo": "NousResearch/Nous-Hermes-2-Yi-34B", "creator": "NousResearch", "baseArchitecture": "01.AI Yi (34B)", "paramsBillions": 34, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-12-23", "lastVerified": "2026-11-01", "ugiScore": 72, "complianceRate": 99.1, "mmluPro": 65.2, "eqBench": 82.9, "ifeval": 81.5, "hfDownloads30d": 9227, "hfLikes": 266, "ollamaCommand": "NousResearch/Nous-Hermes-2-Yi-34B", "vramQ4GB": 20, "vramQ8GB": 38, "vramFP16GB": 74, "summary": "Uncensored DPO/SFT fine-tune of 01.AI Yi (34B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~20 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 34B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "xannie1-qwen3-8-27b-heretic-abliterated-uncensored-gguf", "name": "Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "hfRepo": "xannie1/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "creator": "xannie1", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-03", "lastVerified": "2026-11-01", "ugiScore": 73.6, "complianceRate": 98.6, "mmluPro": 68.5, "eqBench": 82.8, "ifeval": 84.5, "hfDownloads30d": 13127, "hfLikes": 2, "ollamaCommand": "xannie1/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by xannie1, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "heyujie-qwen3-5-35b-a3b-abliterated-gguf", "name": "Qwen3.5-35B-A3B-abliterated-GGUF", "hfRepo": "HeYujie/Qwen3.5-35B-A3B-abliterated-GGUF", "creator": "HeYujie", "baseArchitecture": "Qwen 3.5 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-02-28", "lastVerified": "2026-11-01", "ugiScore": 74.4, "complianceRate": 97.6, "mmluPro": 69.4, "eqBench": 82.7, "ifeval": 85.3, "hfDownloads30d": 12932, "hfLikes": 12, "ollamaCommand": "HeYujie/Qwen3.5-35B-A3B-abliterated-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.5 (35B) checkpoint by HeYujie, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "azminawwar-huihui-qwen3-5-0-8b-abliterated-claude-mythos-distilled-25k", "name": "Huihui-Qwen3.5-0.8B-abliterated-claude_mythos_distilled_25k", "hfRepo": "azminawwar/Huihui-Qwen3.5-0.8B-abliterated-claude_mythos_distilled_25k", "creator": "azminawwar", "baseArchitecture": "Qwen 3.5 (0.8B)", "paramsBillions": 0.8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-01", "lastVerified": "2026-11-01", "ugiScore": 60.2, "complianceRate": 98.4, "mmluPro": 49.5, "eqBench": 70.4, "ifeval": 70.8, "hfDownloads30d": 12827, "hfLikes": 1, "ollamaCommand": "azminawwar/Huihui-Qwen3.5-0.8B-abliterated-claude_mythos_distilled_25k", "vramQ4GB": 1.6, "vramQ8GB": 2, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Qwen 3.5 (0.8B) checkpoint by azminawwar, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.6 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (0.8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 0.8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "qwen-2.5-72b-instruct-abliterated", "name": "Qwen2.5-72B-Instruct-abliterated", "hfRepo": "huihui-ai/Qwen2.5-72B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 (Dense)", "paramsBillions": 72.7, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-14", "lastVerified": "2025-06-06", "ugiScore": 76.4, "complianceRate": 98.2, "mmluPro": 71.8, "eqBench": 84.6, "ifeval": 86.2, "hfDownloads30d": 8420, "hfLikes": 288, "openRouterSlug": "qwen/qwen-2.5-72b-instruct", "ollamaCommand": "huihui-ai/Qwen2.5-72B-Instruct-abliterated", "vramQ4GB": 42, "vramQ8GB": 76, "vramFP16GB": 146, "summary": "Orthogonalized refusal directions removed from Qwen2.5-72B with near-zero degradation in STEM, coding, or 128k long-context recall.", "strengths": [ "98.2% adversarial compliance", "Top-tier coding & STEM retention", "128k native context window" ], "tradeoffs": [ "Requires dual 24GB GPUs or 64GB Unified Memory at Q4_K_M" ] }, { "id": "abiray-huihui-qwen3-5-9b-abliterated-gguf", "name": "Huihui-Qwen3.5-9B-abliterated-GGUF", "hfRepo": "Abiray/Huihui-Qwen3.5-9B-abliterated-GGUF", "creator": "Abiray", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-09", "lastVerified": "2026-11-01", "ugiScore": 68.4, "complianceRate": 98.4, "mmluPro": 61.4, "eqBench": 78, "ifeval": 79.2, "hfDownloads30d": 12620, "hfLikes": 5, "ollamaCommand": "Abiray/Huihui-Qwen3.5-9B-abliterated-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by Abiray, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "firworks-cydonia-24b-v4-3-heretic-nvfp4", "name": "Cydonia-24B-v4.3-heretic-nvfp4", "hfRepo": "Firworks/Cydonia-24B-v4.3-heretic-nvfp4", "creator": "Firworks", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-12-31", "lastVerified": "2026-11-01", "ugiScore": 71, "complianceRate": 99.3, "mmluPro": 63.2, "eqBench": 81.4, "ifeval": 81.4, "hfDownloads30d": 12671, "hfLikes": 1, "ollamaCommand": "Firworks/Cydonia-24B-v4.3-heretic-nvfp4", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Orthogonal weight-projected Mistral Small 3 (24B) checkpoint by Firworks, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB VRAM at Q4_K_M)", "Preserves native Mistral Small 3 (24B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 24B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "doctor-shotgun-ms3-2-24b-magnum-diamond-gguf", "name": "MS3.2-24B-Magnum-Diamond-GGUF", "hfRepo": "Doctor-Shotgun/MS3.2-24B-Magnum-Diamond-GGUF", "creator": "Doctor-Shotgun", "baseArchitecture": "Open Transformer (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-06-22", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 99.6, "mmluPro": 63.9, "eqBench": 80.6, "ifeval": 82, "hfDownloads30d": 12106, "hfLikes": 28, "ollamaCommand": "Doctor-Shotgun/MS3.2-24B-Magnum-Diamond-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (24B) by Doctor-Shotgun, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 24B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "sakamakismile-huihui-qwen3-8-27b-abliterated-nvfp4", "name": "Huihui-Qwen3.8-27B-abliterated-NVFP4", "hfRepo": "sakamakismile/Huihui-Qwen3.8-27B-abliterated-NVFP4", "creator": "sakamakismile", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-16", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 98, "mmluPro": 68.1, "eqBench": 81.5, "ifeval": 83.3, "hfDownloads30d": 11412, "hfLikes": 66, "ollamaCommand": "sakamakismile/Huihui-Qwen3.8-27B-abliterated-NVFP4", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by sakamakismile, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davidau-llama-3-2-8x4b-moe-v2-dark-champion-instruct-uncensored-abliterated-21b-gguf", "name": "Llama-3.2-8X4B-MOE-V2-Dark-Champion-Instruct-uncensored-abliterated-21B-GGUF", "hfRepo": "DavidAU/Llama-3.2-8X4B-MOE-V2-Dark-Champion-Instruct-uncensored-abliterated-21B-GGUF", "creator": "DavidAU", "baseArchitecture": "Llama 3.2 (21B)", "paramsBillions": 21, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-02-12", "lastVerified": "2026-11-01", "ugiScore": 70.9, "complianceRate": 99.2, "mmluPro": 63.9, "eqBench": 81, "ifeval": 81.2, "hfDownloads30d": 9873, "hfLikes": 163, "ollamaCommand": "DavidAU/Llama-3.2-8X4B-MOE-V2-Dark-Champion-Instruct-uncensored-abliterated-21B-GGUF", "vramQ4GB": 13, "vramQ8GB": 24, "vramFP16GB": 46, "summary": "Orthogonal weight-projected Llama 3.2 (21B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~13 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (21B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 21B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-6-mixtral-8x7b", "name": "dolphin-2.6-mixtral-8x7b", "hfRepo": "dphn/dolphin-2.6-mixtral-8x7b", "creator": "dphn", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-12-21", "lastVerified": "2026-11-01", "ugiScore": 74.7, "complianceRate": 99, "mmluPro": 68.1, "eqBench": 84.1, "ifeval": 84.3, "hfDownloads30d": 8957, "hfLikes": 217, "ollamaCommand": "dphn/dolphin-2.6-mixtral-8x7b", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "abiray-huihui-qwen3-6-35b-a3b-abliterated-gguf", "name": "Huihui-Qwen3.6-35B-A3B-abliterated-GGUF", "hfRepo": "Abiray/Huihui-Qwen3.6-35B-A3B-abliterated-GGUF", "creator": "Abiray", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-18", "lastVerified": "2026-11-01", "ugiScore": 74.5, "complianceRate": 98.9, "mmluPro": 68.9, "eqBench": 82.5, "ifeval": 85.4, "hfDownloads30d": 11595, "hfLikes": 19, "ollamaCommand": "Abiray/Huihui-Qwen3.6-35B-A3B-abliterated-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by Abiray, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-l3-8b-stheno-v3-2", "name": "L3-8B-Stheno-v3.2", "hfRepo": "Sao10K/L3-8B-Stheno-v3.2", "creator": "Sao10K", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-06-05", "lastVerified": "2024-06-07", "ugiScore": 64.6, "complianceRate": 98.5, "mmluPro": 55.2, "eqBench": 78.9, "ifeval": 74.8, "hfDownloads30d": 4498, "hfLikes": 481, "ollamaCommand": "Sao10K/L3-8B-Stheno-v3.2", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (8B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "prithivmlmods-qwen3-8-27b-abliterated-gguf", "name": "Qwen3.8-27B-abliterated-GGUF", "hfRepo": "prithivMLmods/Qwen3.8-27B-abliterated-GGUF", "creator": "prithivMLmods", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-16", "lastVerified": "2026-11-01", "ugiScore": 71.6, "complianceRate": 98.4, "mmluPro": 68.5, "eqBench": 81.8, "ifeval": 82.5, "hfDownloads30d": 11523, "hfLikes": 7, "ollamaCommand": "prithivMLmods/Qwen3.8-27B-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by prithivMLmods, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "failspy-llama-3-70b-instruct-abliterated", "name": "llama-3-70B-Instruct-abliterated", "hfRepo": "failspy/llama-3-70B-Instruct-abliterated", "creator": "failspy", "baseArchitecture": "Llama 3 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-07", "lastVerified": "2024-05-07", "ugiScore": 76.2, "complianceRate": 99, "mmluPro": 71.9, "eqBench": 85.8, "ifeval": 86.7, "hfDownloads30d": 8919, "hfLikes": 154, "ollamaCommand": "failspy/llama-3-70B-Instruct-abliterated", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Orthogonal weight-projected Llama 3 (70B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~40 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (70B) instruction and coding schemas" ], "tradeoffs": [ "Requires 40 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-vl-4b-instruct-abliterated", "name": "Huihui-Qwen3-VL-4B-Instruct-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-10-16", "lastVerified": "2026-11-01", "ugiScore": 64.8, "complianceRate": 98.5, "mmluPro": 55.6, "eqBench": 73, "ifeval": 75.5, "hfDownloads30d": 9415, "hfLikes": 120, "ollamaCommand": "huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-1-mixtral-1x22b", "name": "dolphin-2.9.1-mixtral-1x22b", "hfRepo": "dphn/dolphin-2.9.1-mixtral-1x22b", "creator": "dphn", "baseArchitecture": "Mixtral MoE (17.6B)", "paramsBillions": 17.6, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-22", "lastVerified": "2026-11-01", "ugiScore": 68.9, "complianceRate": 98.2, "mmluPro": 61.6, "eqBench": 80.6, "ifeval": 78.3, "hfDownloads30d": 10469, "hfLikes": 46, "ollamaCommand": "dphn/dolphin-2.9.1-mixtral-1x22b", "vramQ4GB": 11, "vramQ8GB": 20, "vramFP16GB": 39, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (17.6B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~11 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 17.6B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "richardyoung-qwen2-5-7b-instruct-abliterated-gguf", "name": "Qwen2.5-7B-Instruct-abliterated-GGUF", "hfRepo": "richardyoung/Qwen2.5-7B-Instruct-abliterated-GGUF", "creator": "richardyoung", "baseArchitecture": "Qwen 2.5 (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-12-01", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 98.8, "mmluPro": 59, "eqBench": 77, "ifeval": 77.8, "hfDownloads30d": 11008, "hfLikes": 6, "ollamaCommand": "richardyoung/Qwen2.5-7B-Instruct-abliterated-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 (7B) checkpoint by richardyoung, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-12b-agentic-fable5-abliterated-gguf", "name": "Huihui-gemma-4-12B-agentic-fable5-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-gemma-4-12B-agentic-fable5-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-22", "lastVerified": "2026-06-22", "ugiScore": 67.9, "complianceRate": 98.2, "mmluPro": 58.8, "eqBench": 78.5, "ifeval": 78.2, "hfDownloads30d": 10337, "hfLikes": 49, "ollamaCommand": "huihui-ai/Huihui-gemma-4-12B-agentic-fable5-abliterated-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 2 (12B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "lord-h4d3zs-qwen3-8-distill-35b-a3b-coder-abliterated", "name": "Qwen3.8-Distill-35B-A3B-Coder-Abliterated", "hfRepo": "Lord-H4D3ZS/Qwen3.8-Distill-35B-A3B-Coder-Abliterated", "creator": "Lord-H4D3ZS", "baseArchitecture": "Qwen 3.8 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-15", "lastVerified": "2026-11-01", "ugiScore": 72.8, "complianceRate": 99, "mmluPro": 72.4, "eqBench": 79.7, "ifeval": 85.9, "hfDownloads30d": 10556, "hfLikes": 25, "ollamaCommand": "Lord-H4D3ZS/Qwen3.8-Distill-35B-A3B-Coder-Abliterated", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.8 (35B) checkpoint by Lord-H4D3ZS, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "hotdogs-qwen3-8-27b-abliterated-mtp-gguf", "name": "Qwen3.8-27B-abliterated-MTP-GGUF", "hfRepo": "hotdogs/Qwen3.8-27B-abliterated-MTP-GGUF", "creator": "hotdogs", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-15", "lastVerified": "2026-11-01", "ugiScore": 72, "complianceRate": 98.8, "mmluPro": 68.5, "eqBench": 81.3, "ifeval": 82.9, "hfDownloads30d": 10592, "hfLikes": 21, "ollamaCommand": "hotdogs/Qwen3.8-27B-abliterated-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by hotdogs, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-pro-mistral-7b", "name": "Hermes-2-Pro-Mistral-7B", "hfRepo": "NousResearch/Hermes-2-Pro-Mistral-7B", "creator": "NousResearch", "baseArchitecture": "Mistral (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-03-11", "lastVerified": "2024-09-08", "ugiScore": 66.6, "complianceRate": 98.5, "mmluPro": 54.1, "eqBench": 75.2, "ifeval": 75.9, "hfDownloads30d": 3209, "hfLikes": 509, "ollamaCommand": "NousResearch/Hermes-2-Pro-Mistral-7B", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Mistral (7B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "therickcain-qwen3-8-27b-heretic-abliterated-uncensored-gguf", "name": "Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "hfRepo": "therickcain/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "creator": "therickcain", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-29", "lastVerified": "2026-11-01", "ugiScore": 71.5, "complianceRate": 97.7, "mmluPro": 67.4, "eqBench": 82.8, "ifeval": 82.4, "hfDownloads30d": 10823, "hfLikes": 1, "ollamaCommand": "therickcain/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by therickcain, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-rocinante-12b-v1-1-gguf", "name": "Rocinante-12B-v1.1-GGUF", "hfRepo": "TheDrummer/Rocinante-12B-v1.1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-08-15", "lastVerified": "2026-11-01", "ugiScore": 68, "complianceRate": 99.1, "mmluPro": 56.9, "eqBench": 82.5, "ifeval": 78.3, "hfDownloads30d": 9711, "hfLikes": 69, "ollamaCommand": "TheDrummer/Rocinante-12B-v1.1-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "windowsxp811203-qwen3-8-27b-abliterated-gguf", "name": "Qwen3.8-27B-Abliterated-GGUF", "hfRepo": "windowsxp811203/Qwen3.8-27B-Abliterated-GGUF", "creator": "windowsxp811203", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-15", "lastVerified": "2026-11-01", "ugiScore": 73.4, "complianceRate": 99, "mmluPro": 67.1, "eqBench": 82.1, "ifeval": 84.3, "hfDownloads30d": 10602, "hfLikes": 9, "ollamaCommand": "windowsxp811203/Qwen3.8-27B-Abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by windowsxp811203, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "barding-defense-qwen3-8-27b-huihui-abliterated-groupwise-int-ninfer", "name": "Qwen3.8-27B-huihui-abliterated-groupwise-int-NInfer", "hfRepo": "Barding-Defense/Qwen3.8-27B-huihui-abliterated-groupwise-int-NInfer", "creator": "Barding-Defense", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-03", "lastVerified": "2026-11-01", "ugiScore": 73, "complianceRate": 98, "mmluPro": 68.8, "eqBench": 82.1, "ifeval": 83.9, "hfDownloads30d": 10662, "hfLikes": 1, "ollamaCommand": "Barding-Defense/Qwen3.8-27B-huihui-abliterated-groupwise-int-NInfer", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by Barding-Defense, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "vicgalle-configurable-hermes-2-pro-llama-3-8b", "name": "Configurable-Hermes-2-Pro-Llama-3-8B", "hfRepo": "vicgalle/Configurable-Hermes-2-Pro-Llama-3-8B", "creator": "vicgalle", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-02", "lastVerified": "2026-11-01", "ugiScore": 65.9, "complianceRate": 98.8, "mmluPro": 55.6, "eqBench": 77.2, "ifeval": 75.2, "hfDownloads30d": 10415, "hfLikes": 6, "ollamaCommand": "vicgalle/Configurable-Hermes-2-Pro-Llama-3-8B", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by vicgalle, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "mlabonne-daredevil-8b-abliterated", "name": "Daredevil-8B-abliterated", "hfRepo": "mlabonne/Daredevil-8B-abliterated", "creator": "mlabonne", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-26", "lastVerified": "2026-01-22", "ugiScore": 65.2, "complianceRate": 98.5, "mmluPro": 57.3, "eqBench": 77.1, "ifeval": 75.4, "hfDownloads30d": 9062, "hfLikes": 63, "ollamaCommand": "mlabonne/Daredevil-8B-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Open Transformer (8B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-pro-llama-3-8b", "name": "Hermes-2-Pro-Llama-3-8B", "hfRepo": "NousResearch/Hermes-2-Pro-Llama-3-8B", "creator": "NousResearch", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-04-30", "lastVerified": "2024-09-14", "ugiScore": 65.6, "complianceRate": 99.3, "mmluPro": 56.5, "eqBench": 77.2, "ifeval": 74.9, "hfDownloads30d": 3147, "hfLikes": 457, "ollamaCommand": "NousResearch/Hermes-2-Pro-Llama-3-8B", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "failspy-meta-llama-3-70b-instruct-abliterated-v3-5", "name": "Meta-Llama-3-70B-Instruct-abliterated-v3.5", "hfRepo": "failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5", "creator": "failspy", "baseArchitecture": "Llama 3 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-28", "lastVerified": "2024-05-30", "ugiScore": 75.5, "complianceRate": 98.3, "mmluPro": 71.2, "eqBench": 86, "ifeval": 86, "hfDownloads30d": 9139, "hfLikes": 56, "ollamaCommand": "failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Orthogonal weight-projected Llama 3 (70B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~40 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (70B) instruction and coding schemas" ], "tradeoffs": [ "Requires 40 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-3-mistral-nemo-12b-gguf", "name": "dolphin-2.9.3-mistral-nemo-12b-gguf", "hfRepo": "dphn/dolphin-2.9.3-mistral-nemo-12b-gguf", "creator": "dphn", "baseArchitecture": "Mistral NeMo (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-07-24", "lastVerified": "2026-11-01", "ugiScore": 67.6, "complianceRate": 99.5, "mmluPro": 57.9, "eqBench": 77.6, "ifeval": 77, "hfDownloads30d": 8875, "hfLikes": 73, "ollamaCommand": "dphn/dolphin-2.9.3-mistral-nemo-12b-gguf", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Mistral NeMo (12B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "maerong3-huihui-mistral-small-3-2-24b-instruct-2506-abliterated-q4-k-m-gguf", "name": "Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated-Q4_K_M-GGUF", "hfRepo": "maerong3/Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated-Q4_K_M-GGUF", "creator": "maerong3", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-08-06", "lastVerified": "2026-11-01", "ugiScore": 69.8, "complianceRate": 99.3, "mmluPro": 64, "eqBench": 80.4, "ifeval": 80.2, "hfDownloads30d": 9855, "hfLikes": 7, "ollamaCommand": "maerong3/Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated-Q4_K_M-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Orthogonal weight-projected Mistral Small 3 (24B) checkpoint by maerong3, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB VRAM at Q4_K_M)", "Preserves native Mistral Small 3 (24B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 24B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-1-llama-3-70b", "name": "dolphin-2.9.1-llama-3-70b", "hfRepo": "dphn/dolphin-2.9.1-llama-3-70b", "creator": "dphn", "baseArchitecture": "Llama 3 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-22", "lastVerified": "2026-11-01", "ugiScore": 75.4, "complianceRate": 98.4, "mmluPro": 70.4, "eqBench": 86.8, "ifeval": 85, "hfDownloads30d": 8943, "hfLikes": 62, "ollamaCommand": "dphn/dolphin-2.9.1-llama-3-70b", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (70B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (40 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "noteventhrice-deepseek-r1-0528-qwen3-8b-abliterated-q4-k-m-gguf", "name": "DeepSeek-R1-0528-Qwen3-8B-abliterated-Q4_K_M-GGUF", "hfRepo": "noteventhrice/DeepSeek-R1-0528-Qwen3-8B-abliterated-Q4_K_M-GGUF", "creator": "noteventhrice", "baseArchitecture": "DeepSeek R1 Distill Qwen (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-05-30", "lastVerified": "2026-11-01", "ugiScore": 67.9, "complianceRate": 98.4, "mmluPro": 60.8, "eqBench": 75.9, "ifeval": 78.6, "hfDownloads30d": 9853, "hfLikes": 1, "ollamaCommand": "noteventhrice/DeepSeek-R1-0528-Qwen3-8B-abliterated-Q4_K_M-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (8B) checkpoint by noteventhrice, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "phillylabs-kat-coder-v2-5-dev-35b-a3b-abliterated-uncensored-philadelphia-class", "name": "KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS", "hfRepo": "PhillyLabs/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS", "creator": "PhillyLabs", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-07", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 98.4, "mmluPro": 68.7, "eqBench": 80.7, "ifeval": 85, "hfDownloads30d": 9743, "hfLikes": 6, "ollamaCommand": "PhillyLabs/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Open Transformer (35B) checkpoint by PhillyLabs, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-1-llama-3-8b", "name": "dolphin-2.9.1-llama-3-8b", "hfRepo": "dphn/dolphin-2.9.1-llama-3-8b", "creator": "dphn", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-10", "lastVerified": "2026-11-01", "ugiScore": 66.4, "complianceRate": 99.3, "mmluPro": 55.9, "eqBench": 75.8, "ifeval": 75.7, "hfDownloads30d": 9158, "hfLikes": 38, "ollamaCommand": "dphn/dolphin-2.9.1-llama-3-8b", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "prithivmlmods-gliese-qwen3-5-9b-abliterated-caption", "name": "Gliese-Qwen3.5-9B-Abliterated-Caption", "hfRepo": "prithivMLmods/Gliese-Qwen3.5-9B-Abliterated-Caption", "creator": "prithivMLmods", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-10", "lastVerified": "2026-11-01", "ugiScore": 66.4, "complianceRate": 97.6, "mmluPro": 60.2, "eqBench": 78.1, "ifeval": 77.2, "hfDownloads30d": 8737, "hfLikes": 65, "ollamaCommand": "prithivMLmods/Gliese-Qwen3.5-9B-Abliterated-Caption", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by prithivMLmods, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-9b-abliterated", "name": "Huihui-Qwen3.5-9B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-9B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-09", "lastVerified": "2026-03-09", "ugiScore": 67.7, "complianceRate": 98.9, "mmluPro": 60.9, "eqBench": 76.4, "ifeval": 78.5, "hfDownloads30d": 7240, "hfLikes": 163, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-9B-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-meta-llama-3-1-8b-instruct-abliterated", "name": "Meta-Llama-3.1-8B-Instruct-abliterated", "hfRepo": "mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated", "creator": "mlabonne", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-07-24", "lastVerified": "2024-09-14", "ugiScore": 66.6, "complianceRate": 98.4, "mmluPro": 56.9, "eqBench": 78, "ifeval": 78.1, "hfDownloads30d": 6396, "hfLikes": 219, "ollamaCommand": "mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3.1 (8B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.1 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-2-mixtral-8x22b", "name": "dolphin-2.9.2-mixtral-8x22b", "hfRepo": "dphn/dolphin-2.9.2-mixtral-8x22b", "creator": "dphn", "baseArchitecture": "Mixtral MoE (140.8B)", "paramsBillions": 140.8, "parameterClass": "Frontier (100B+)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-28", "lastVerified": "2026-11-01", "ugiScore": 76.5, "complianceRate": 99.4, "mmluPro": 70.3, "eqBench": 85.1, "ifeval": 86.1, "hfDownloads30d": 9010, "hfLikes": 44, "ollamaCommand": "dphn/dolphin-2.9.2-mixtral-8x22b", "vramQ4GB": 80, "vramQ8GB": 154, "vramFP16GB": 302, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (140.8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on multi-GPU workstations or cloud nodes (~80 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (80 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "boldingbuilds-nvidia-nemotron-3-5-bolt-74b-a9b-abliterated-24gb-gguf", "name": "NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Abliterated-24GB-GGUF", "hfRepo": "BoldingBuilds/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Abliterated-24GB-GGUF", "creator": "BoldingBuilds", "baseArchitecture": "NVIDIA Nemotron (74B)", "paramsBillions": 74, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "NVIDIA Open Model", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-16", "lastVerified": "2026-11-01", "ugiScore": 74.7, "complianceRate": 98.6, "mmluPro": 71.6, "eqBench": 86.3, "ifeval": 85.2, "hfDownloads30d": 9245, "hfLikes": 12, "ollamaCommand": "BoldingBuilds/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Abliterated-24GB-GGUF", "vramQ4GB": 43, "vramQ8GB": 81, "vramFP16GB": 159, "summary": "Orthogonal weight-projected NVIDIA Nemotron (74B) checkpoint by BoldingBuilds, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~43 GB VRAM at Q4_K_M)", "Preserves native NVIDIA Nemotron (74B) instruction and coding schemas" ], "tradeoffs": [ "Requires 43 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "dphn-dolphin3-0-llama3-1-8b", "name": "Dolphin3.0-Llama3.1-8B", "hfRepo": "dphn/Dolphin3.0-Llama3.1-8B", "creator": "dphn", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-12-29", "lastVerified": "2026-11-01", "ugiScore": 67.4, "complianceRate": 98.7, "mmluPro": 56.5, "eqBench": 77, "ifeval": 76.7, "hfDownloads30d": 3210, "hfLikes": 414, "ollamaCommand": "dphn/Dolphin3.0-Llama3.1-8B", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "ayozhee-qwen3-8-27b-abliterated-gsq-orca-gguf", "name": "Qwen3.8-27B-Abliterated-GSQ-Orca-GGUF", "hfRepo": "ayozhee/Qwen3.8-27B-Abliterated-GSQ-Orca-GGUF", "creator": "ayozhee", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-08", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 98.9, "mmluPro": 67.6, "eqBench": 82.6, "ifeval": 83.6, "hfDownloads30d": 8680, "hfLikes": 41, "ollamaCommand": "ayozhee/Qwen3.8-27B-Abliterated-GSQ-Orca-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by ayozhee, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-1-yi-1-5-9b", "name": "dolphin-2.9.1-yi-1.5-9b", "hfRepo": "dphn/dolphin-2.9.1-yi-1.5-9b", "creator": "dphn", "baseArchitecture": "01.AI Yi (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-18", "lastVerified": "2026-11-01", "ugiScore": 66.2, "complianceRate": 99.4, "mmluPro": 56.4, "eqBench": 77.3, "ifeval": 75.6, "hfDownloads30d": 8902, "hfLikes": 25, "ollamaCommand": "dphn/dolphin-2.9.1-yi-1.5-9b", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Uncensored DPO/SFT fine-tune of 01.AI Yi (9B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~6.2 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 9B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "orenguteng-llama-3-1-8b-lexi-uncensored-v2-gguf", "name": "Llama-3.1-8B-Lexi-Uncensored-V2-GGUF", "hfRepo": "Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2-GGUF", "creator": "Orenguteng", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-09", "lastVerified": "2026-11-01", "ugiScore": 67.9, "complianceRate": 98.5, "mmluPro": 56.6, "eqBench": 78.7, "ifeval": 78.5, "hfDownloads30d": 5611, "hfLikes": 238, "ollamaCommand": "Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by Orenguteng, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-3-yi-1-5-34b-32k", "name": "dolphin-2.9.3-Yi-1.5-34B-32k", "hfRepo": "dphn/dolphin-2.9.3-Yi-1.5-34B-32k", "creator": "dphn", "baseArchitecture": "01.AI Yi (34B)", "paramsBillions": 34, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-06-23", "lastVerified": "2026-11-01", "ugiScore": 72.2, "complianceRate": 99.3, "mmluPro": 65.5, "eqBench": 82.9, "ifeval": 81.7, "hfDownloads30d": 8860, "hfLikes": 17, "ollamaCommand": "dphn/dolphin-2.9.3-Yi-1.5-34B-32k", "vramQ4GB": 20, "vramQ8GB": 38, "vramFP16GB": 74, "summary": "Uncensored DPO/SFT fine-tune of 01.AI Yi (34B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~20 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 34B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "failspy-smaug-llama-3-70b-instruct-abliterated-v3", "name": "Smaug-Llama-3-70B-Instruct-abliterated-v3", "hfRepo": "failspy/Smaug-Llama-3-70B-Instruct-abliterated-v3", "creator": "failspy", "baseArchitecture": "Llama 3 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-20", "lastVerified": "2024-05-22", "ugiScore": 74.4, "complianceRate": 97.8, "mmluPro": 71.4, "eqBench": 86.8, "ifeval": 84.9, "hfDownloads30d": 8850, "hfLikes": 16, "ollamaCommand": "failspy/Smaug-Llama-3-70B-Instruct-abliterated-v3", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Orthogonal weight-projected Llama 3 (70B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~40 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (70B) instruction and coding schemas" ], "tradeoffs": [ "Requires 40 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "weyaxi-bagel-hermes-34b-slerp", "name": "Bagel-Hermes-34B-Slerp", "hfRepo": "Weyaxi/Bagel-Hermes-34B-Slerp", "creator": "Weyaxi", "baseArchitecture": "Llama 3.1 (34B)", "paramsBillions": 34, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-12", "lastVerified": "2026-11-01", "ugiScore": 75.1, "complianceRate": 98.3, "mmluPro": 64.6, "eqBench": 84, "ifeval": 85.9, "hfDownloads30d": 8955, "hfLikes": 1, "ollamaCommand": "Weyaxi/Bagel-Hermes-34B-Slerp", "vramQ4GB": 20, "vramQ8GB": 38, "vramFP16GB": 74, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (34B) by Weyaxi, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~20 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 34B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-rocinante-xl-16b-v1-gguf", "name": "Rocinante-XL-16B-v1-GGUF", "hfRepo": "TheDrummer/Rocinante-XL-16B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (16B)", "paramsBillions": 16, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-04-18", "lastVerified": "2026-08-05", "ugiScore": 67.5, "complianceRate": 98.9, "mmluPro": 60.2, "eqBench": 83.4, "ifeval": 77.9, "hfDownloads30d": 8359, "hfLikes": 39, "ollamaCommand": "TheDrummer/Rocinante-XL-16B-v1-GGUF", "vramQ4GB": 10, "vramQ8GB": 19, "vramFP16GB": 35, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (16B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~10 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "failspy-codestral-22b-v0-1-abliterated-v3", "name": "Codestral-22B-v0.1-abliterated-v3", "hfRepo": "failspy/Codestral-22B-v0.1-abliterated-v3", "creator": "failspy", "baseArchitecture": "Open Transformer (22B)", "paramsBillions": 22, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-06-03", "lastVerified": "2024-06-03", "ugiScore": 71.1, "complianceRate": 98.6, "mmluPro": 63.1, "eqBench": 81.9, "ifeval": 81.5, "hfDownloads30d": 8726, "hfLikes": 13, "ollamaCommand": "failspy/Codestral-22B-v0.1-abliterated-v3", "vramQ4GB": 14, "vramQ8GB": 25, "vramFP16GB": 48, "summary": "Orthogonal weight-projected Open Transformer (22B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~14 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (22B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 22B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "weyaxi-nous-hermes-2-sus-chat-34b-slerp", "name": "Nous-Hermes-2-SUS-Chat-34B-Slerp", "hfRepo": "Weyaxi/Nous-Hermes-2-SUS-Chat-34B-Slerp", "creator": "Weyaxi", "baseArchitecture": "Open Transformer (34B)", "paramsBillions": 34, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-02", "lastVerified": "2026-11-01", "ugiScore": 73.2, "complianceRate": 98.7, "mmluPro": 65.5, "eqBench": 83.2, "ifeval": 82.7, "hfDownloads30d": 8890, "hfLikes": 1, "ollamaCommand": "Weyaxi/Nous-Hermes-2-SUS-Chat-34B-Slerp", "vramQ4GB": 20, "vramQ8GB": 38, "vramFP16GB": 74, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (34B) by Weyaxi, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~20 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 34B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "fyb1214-huihui-qwen3-8-27b-abliterated-ternary-bonsai-2-ninfer", "name": "Huihui-Qwen3.8-27B-Abliterated-Ternary-Bonsai-2-NInfer", "hfRepo": "fyb1214/Huihui-Qwen3.8-27B-Abliterated-Ternary-Bonsai-2-NInfer", "creator": "fyb1214", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-04", "lastVerified": "2026-11-01", "ugiScore": 73.7, "complianceRate": 98.1, "mmluPro": 67.9, "eqBench": 82.5, "ifeval": 84.6, "hfDownloads30d": 8821, "hfLikes": 2, "ollamaCommand": "fyb1214/Huihui-Qwen3.8-27B-Abliterated-Ternary-Bonsai-2-NInfer", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by fyb1214, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui4-8b-a4b-gguf", "name": "Huihui4-8B-A4B-GGUF", "hfRepo": "huihui-ai/Huihui4-8B-A4B-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-25", "lastVerified": "2026-04-25", "ugiScore": 64.7, "complianceRate": 99.2, "mmluPro": 56.1, "eqBench": 76, "ifeval": 74.9, "hfDownloads30d": 8626, "hfLikes": 13, "ollamaCommand": "huihui-ai/Huihui4-8B-A4B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Open Transformer (8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-4-14b", "name": "Hermes-4-14B", "hfRepo": "NousResearch/Hermes-4-14B", "creator": "NousResearch", "baseArchitecture": "Open Transformer (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-08-30", "lastVerified": "2026-01-09", "ugiScore": 68.8, "complianceRate": 98.4, "mmluPro": 59.5, "eqBench": 79.6, "ifeval": 78.2, "hfDownloads30d": 5509, "hfLikes": 198, "ollamaCommand": "NousResearch/Hermes-4-14B", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (14B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 14B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-3-llama-3-2-3b-gguf", "name": "Hermes-3-Llama-3.2-3B-GGUF", "hfRepo": "NousResearch/Hermes-3-Llama-3.2-3B-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3.2 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-12-11", "lastVerified": "2024-12-18", "ugiScore": 64, "complianceRate": 98.3, "mmluPro": 49, "eqBench": 75.7, "ifeval": 74.5, "hfDownloads30d": 7351, "hfLikes": 75, "ollamaCommand": "NousResearch/Hermes-3-Llama-3.2-3B-GGUF", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.2 (3B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.9 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 3B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-e2b-it-qat-q4-0-unquantized-abliterated-gguf", "name": "Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-09", "lastVerified": "2026-06-11", "ugiScore": 59.1, "complianceRate": 99.3, "mmluPro": 47.8, "eqBench": 71.7, "ifeval": 69.1, "hfDownloads30d": 8116, "hfLikes": 16, "ollamaCommand": "huihui-ai/Huihui-gemma-4-E2B-it-qat-q4_0-unquantized-abliterated-GGUF", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Orthogonal weight-projected Gemma 2 (2B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (2B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 2B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "artomyuan-qwen3-8-27b-coder390-efficientthink-abliterated-rocmfpx", "name": "Qwen3.8-27B-Coder390-EfficientThink-abliterated-ROCmFPX", "hfRepo": "ArtomYuan/Qwen3.8-27B-Coder390-EfficientThink-abliterated-ROCmFPX", "creator": "ArtomYuan", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-14", "lastVerified": "2026-11-01", "ugiScore": 73.2, "complianceRate": 98.8, "mmluPro": 70.2, "eqBench": 79.3, "ifeval": 86.3, "hfDownloads30d": 8245, "hfLikes": 1, "ollamaCommand": "ArtomYuan/Qwen3.8-27B-Coder390-EfficientThink-abliterated-ROCmFPX", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by ArtomYuan, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-27b-it-abliterated", "name": "gemma-3-27b-it-abliterated", "hfRepo": "mlabonne/gemma-3-27b-it-abliterated", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-03-16", "lastVerified": "2025-03-21", "ugiScore": 70.9, "complianceRate": 99.3, "mmluPro": 65.2, "eqBench": 84.4, "ifeval": 81.3, "hfDownloads30d": 2989, "hfLikes": 347, "ollamaCommand": "mlabonne/gemma-3-27b-it-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Gemma 3 (27B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-35b-a3b-abliterated", "name": "Huihui-Qwen3.5-35B-A3B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-35B-A3B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-02-27", "lastVerified": "2026-03-02", "ugiScore": 73.1, "complianceRate": 98.1, "mmluPro": 69.1, "eqBench": 82.4, "ifeval": 84, "hfDownloads30d": 2022, "hfLikes": 410, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-35B-A3B-abliterated", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.5 (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-3-llama-3-2-3b", "name": "Hermes-3-Llama-3.2-3B", "hfRepo": "NousResearch/Hermes-3-Llama-3.2-3B", "creator": "NousResearch", "baseArchitecture": "Llama 3.2 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-12-03", "lastVerified": "2024-12-18", "ugiScore": 63.8, "complianceRate": 98.1, "mmluPro": 48.7, "eqBench": 74.4, "ifeval": 74.3, "hfDownloads30d": 5281, "hfLikes": 184, "ollamaCommand": "NousResearch/Hermes-3-Llama-3.2-3B", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.2 (3B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.9 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 3B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-skyfall-31b-v4-2-gguf", "name": "Skyfall-31B-v4.2-GGUF", "hfRepo": "TheDrummer/Skyfall-31B-v4.2-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-02-13", "lastVerified": "2026-04-03", "ugiScore": 70.8, "complianceRate": 98.9, "mmluPro": 63.5, "eqBench": 86.1, "ifeval": 81.2, "hfDownloads30d": 7319, "hfLikes": 48, "ollamaCommand": "TheDrummer/Skyfall-31B-v4.2-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (31B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "gcsa-ailab-qwen3-8-27b-abliterated-mtp-gguf", "name": "Qwen3.8-27B-Abliterated-MTP-GGUF", "hfRepo": "GCSA-AiLab/Qwen3.8-27B-Abliterated-MTP-GGUF", "creator": "GCSA-AiLab", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-19", "lastVerified": "2026-11-01", "ugiScore": 72, "complianceRate": 98.2, "mmluPro": 67.2, "eqBench": 81.7, "ifeval": 82.9, "hfDownloads30d": 8023, "hfLikes": 1, "ollamaCommand": "GCSA-AiLab/Qwen3.8-27B-Abliterated-MTP-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by GCSA-AiLab, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "lyf-qwen3-8-27b-huihui-abliterated-ninfer-nvfp4", "name": "Qwen3.8-27B-Huihui-Abliterated-NInfer-NVFP4", "hfRepo": "lyf/Qwen3.8-27B-Huihui-Abliterated-NInfer-NVFP4", "creator": "lyf", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-19", "lastVerified": "2026-11-01", "ugiScore": 73.6, "complianceRate": 99.2, "mmluPro": 68.5, "eqBench": 81.3, "ifeval": 84.5, "hfDownloads30d": 7742, "hfLikes": 19, "ollamaCommand": "lyf/Qwen3.8-27B-Huihui-Abliterated-NInfer-NVFP4", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by lyf, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-neohorse-1-4b-abliterated-gguf", "name": "Huihui-NeoHorse-1-4B-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-NeoHorse-1-4B-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-18", "lastVerified": "2026-09-18", "ugiScore": 61.5, "complianceRate": 98, "mmluPro": 52.9, "eqBench": 74, "ifeval": 71.7, "hfDownloads30d": 7673, "hfLikes": 20, "ollamaCommand": "huihui-ai/Huihui-NeoHorse-1-4B-abliterated-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Open Transformer (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "cs2764-huihui-qwen3-5-35b-a3b-abliterated-q6-k-gguf", "name": "Huihui-Qwen3.5-35B-A3B-abliterated-Q6_K-GGUF", "hfRepo": "cs2764/Huihui-Qwen3.5-35B-A3B-abliterated-Q6_K-GGUF", "creator": "cs2764", "baseArchitecture": "Qwen 3.5 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-02-27", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 98.3, "mmluPro": 70.6, "eqBench": 82.3, "ifeval": 83.6, "hfDownloads30d": 7785, "hfLikes": 1, "ollamaCommand": "cs2764/Huihui-Qwen3.5-35B-A3B-abliterated-Q6_K-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.5 (35B) checkpoint by cs2764, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "gryphe-mythomax-l2-13b", "name": "MythoMax-L2-13b", "hfRepo": "Gryphe/MythoMax-L2-13b", "creator": "Gryphe", "baseArchitecture": "Open Transformer (13B)", "paramsBillions": 13, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2023-08-10", "lastVerified": "2026-11-01", "ugiScore": 67.5, "complianceRate": 99.2, "mmluPro": 57.7, "eqBench": 82.3, "ifeval": 77.8, "hfDownloads30d": 1482, "hfLikes": 411, "ollamaCommand": "Gryphe/MythoMax-L2-13b", "vramQ4GB": 8, "vramQ8GB": 15, "vramFP16GB": 29, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (13B) by Gryphe, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-13b", "name": "Nous-Hermes-13b", "hfRepo": "NousResearch/Nous-Hermes-13b", "creator": "NousResearch", "baseArchitecture": "Open Transformer (13B)", "paramsBillions": 13, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-06-03", "lastVerified": "2026-11-01", "ugiScore": 68.9, "complianceRate": 99.6, "mmluPro": 60, "eqBench": 79, "ifeval": 78.3, "hfDownloads30d": 784, "hfLikes": 456, "ollamaCommand": "NousResearch/Nous-Hermes-13b", "vramQ4GB": 8, "vramQ8GB": 15, "vramFP16GB": 29, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (13B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 13B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-llama2-13b", "name": "Nous-Hermes-Llama2-13b", "hfRepo": "NousResearch/Nous-Hermes-Llama2-13b", "creator": "NousResearch", "baseArchitecture": "Open Transformer (13B)", "paramsBillions": 13, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-07-20", "lastVerified": "2024-04-23", "ugiScore": 68.3, "complianceRate": 99, "mmluPro": 59.6, "eqBench": 78.2, "ifeval": 77.7, "hfDownloads30d": 2783, "hfLikes": 322, "ollamaCommand": "NousResearch/Nous-Hermes-Llama2-13b", "vramQ4GB": 8, "vramQ8GB": 15, "vramFP16GB": 29, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (13B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 13B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-llama3-8b-gguf", "name": "dolphin-2.9-llama3-8b-gguf", "hfRepo": "dphn/dolphin-2.9-llama3-8b-gguf", "creator": "dphn", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-04-21", "lastVerified": "2026-11-01", "ugiScore": 66.2, "complianceRate": 99.1, "mmluPro": 55.8, "eqBench": 76.1, "ifeval": 75.5, "hfDownloads30d": 6003, "hfLikes": 94, "ollamaCommand": "dphn/dolphin-2.9-llama3-8b-gguf", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-qwen2-5-7b-instruct-abliterated-v2", "name": "Qwen2.5-7B-Instruct-abliterated-v2", "hfRepo": "huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-09-22", "lastVerified": "2026-11-01", "ugiScore": 65.4, "complianceRate": 98.3, "mmluPro": 59, "eqBench": 77.1, "ifeval": 76.1, "hfDownloads30d": 6229, "hfLikes": 73, "ollamaCommand": "huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 (7B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "spk1337-huihui-qwen3-8-27b-swift1-5-abliterated-iq3s-dflash2-ninfer", "name": "Huihui-Qwen3.8-27B-Swift1.5-Abliterated-IQ3S-DFlash2-NInfer", "hfRepo": "SPK1337/Huihui-Qwen3.8-27B-Swift1.5-Abliterated-IQ3S-DFlash2-NInfer", "creator": "SPK1337", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-19", "lastVerified": "2026-11-01", "ugiScore": 71.8, "complianceRate": 98, "mmluPro": 67.5, "eqBench": 82.4, "ifeval": 82.7, "hfDownloads30d": 7185, "hfLikes": 4, "ollamaCommand": "SPK1337/Huihui-Qwen3.8-27B-Swift1.5-Abliterated-IQ3S-DFlash2-NInfer", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by SPK1337, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-deephermes-3-llama-3-8b-preview", "name": "DeepHermes-3-Llama-3-8B-Preview", "hfRepo": "NousResearch/DeepHermes-3-Llama-3-8B-Preview", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-02-12", "lastVerified": "2025-04-10", "ugiScore": 67.1, "complianceRate": 99.3, "mmluPro": 54.9, "eqBench": 78.5, "ifeval": 77.7, "hfDownloads30d": 1643, "hfLikes": 373, "ollamaCommand": "NousResearch/DeepHermes-3-Llama-3-8B-Preview", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-orion-26b-a4b-v1", "name": "Orion-26B-A4B-v1", "hfRepo": "TheDrummer/Orion-26B-A4B-v1", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-08-26", "lastVerified": "2026-09-06", "ugiScore": 71, "complianceRate": 99.7, "mmluPro": 63, "eqBench": 84.6, "ifeval": 81.3, "hfDownloads30d": 6556, "hfLikes": 39, "ollamaCommand": "TheDrummer/Orion-26B-A4B-v1", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (26B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-12b-it-abliterated-v2-gguf", "name": "gemma-3-12b-it-abliterated-v2-GGUF", "hfRepo": "mlabonne/gemma-3-12b-it-abliterated-v2-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-05-28", "lastVerified": "2025-05-29", "ugiScore": 67.1, "complianceRate": 98, "mmluPro": 58.6, "eqBench": 80.1, "ifeval": 77.4, "hfDownloads30d": 6114, "hfLikes": 63, "ollamaCommand": "mlabonne/gemma-3-12b-it-abliterated-v2-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 3 (12B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "artomyuan-qwen3-8-27b-abliterated-rocmfpx", "name": "Qwen3.8-27B-abliterated-ROCmFPX", "hfRepo": "ArtomYuan/Qwen3.8-27B-abliterated-ROCmFPX", "creator": "ArtomYuan", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-26", "lastVerified": "2026-11-01", "ugiScore": 72.2, "complianceRate": 99, "mmluPro": 67.7, "eqBench": 82.3, "ifeval": 83.1, "hfDownloads30d": 6955, "hfLikes": 3, "ollamaCommand": "ArtomYuan/Qwen3.8-27B-abliterated-ROCmFPX", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by ArtomYuan, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "jj33512-huihui-qwen3-8-27b-abliterated-gguf", "name": "Huihui-Qwen3.8-27B-abliterated-GGUF", "hfRepo": "jj33512/Huihui-Qwen3.8-27B-abliterated-GGUF", "creator": "jj33512", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-13", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 98.3, "mmluPro": 68.1, "eqBench": 80.9, "ifeval": 83.6, "hfDownloads30d": 6957, "hfLikes": 1, "ollamaCommand": "jj33512/Huihui-Qwen3.8-27B-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by jj33512, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "lumen-models-qwen2-5-7b-instruct-abliterated-v2-q4-k-m-gguf", "name": "Qwen2.5-7B-Instruct-abliterated-v2-Q4_K_M-GGUF", "hfRepo": "lumen-models/Qwen2.5-7B-Instruct-abliterated-v2-Q4_K_M-GGUF", "creator": "lumen-models", "baseArchitecture": "Qwen 2.5 (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-11", "lastVerified": "2026-11-01", "ugiScore": 65.3, "complianceRate": 98.8, "mmluPro": 59.2, "eqBench": 76.5, "ifeval": 76, "hfDownloads30d": 6813, "hfLikes": 10, "ollamaCommand": "lumen-models/Qwen2.5-7B-Instruct-abliterated-v2-Q4_K_M-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 (7B) checkpoint by lumen-models, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-4-3-36b", "name": "Hermes-4.3-36B", "hfRepo": "NousResearch/Hermes-4.3-36B", "creator": "NousResearch", "baseArchitecture": "Open Transformer (36B)", "paramsBillions": 36, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-11-17", "lastVerified": "2025-12-06", "ugiScore": 72.4, "complianceRate": 99.3, "mmluPro": 66.3, "eqBench": 83.8, "ifeval": 82, "hfDownloads30d": 1982, "hfLikes": 312, "ollamaCommand": "NousResearch/Hermes-4.3-36B", "vramQ4GB": 21, "vramQ8GB": 40, "vramFP16GB": 78, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (36B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 36B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v4-12b-gguf", "name": "magnum-v4-12b-gguf", "hfRepo": "anthracite-org/magnum-v4-12b-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-10-09", "lastVerified": "2024-11-25", "ugiScore": 67.2, "complianceRate": 98.3, "mmluPro": 57.9, "eqBench": 79, "ifeval": 76.6, "hfDownloads30d": 5774, "hfLikes": 53, "ollamaCommand": "anthracite-org/magnum-v4-12b-gguf", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (12B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "duoneural-lfm2-5-8b-a1b-hermes-agentic-coder-abliterated-v3-gguf", "name": "LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-GGUF", "hfRepo": "DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-GGUF", "creator": "DuoNeural", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-18", "lastVerified": "2026-11-01", "ugiScore": 64.3, "complianceRate": 98.2, "mmluPro": 59.6, "eqBench": 74.8, "ifeval": 76.7, "hfDownloads30d": 6259, "hfLikes": 11, "ollamaCommand": "DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Open Transformer (8B) checkpoint by DuoNeural, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-magidonia-24b-v4-3-gguf", "name": "Magidonia-24B-v4.3-GGUF", "hfRepo": "TheDrummer/Magidonia-24B-v4.3-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-11-29", "lastVerified": "2025-12-17", "ugiScore": 71.2, "complianceRate": 99.7, "mmluPro": 62, "eqBench": 84.1, "ifeval": 81.6, "hfDownloads30d": 5385, "hfLikes": 62, "ollamaCommand": "TheDrummer/Magidonia-24B-v4.3-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "lukey03-qwen3-5-9b-abliterated", "name": "Qwen3.5-9B-abliterated", "hfRepo": "lukey03/Qwen3.5-9B-abliterated", "creator": "lukey03", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-03", "lastVerified": "2026-11-01", "ugiScore": 67.2, "complianceRate": 98.4, "mmluPro": 59.7, "eqBench": 77, "ifeval": 78, "hfDownloads30d": 3264, "hfLikes": 194, "ollamaCommand": "lukey03/Qwen3.5-9B-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by lukey03, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-4-llama3-1-8b-gguf", "name": "dolphin-2.9.4-llama3.1-8b-gguf", "hfRepo": "dphn/dolphin-2.9.4-llama3.1-8b-gguf", "creator": "dphn", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-07", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 99.2, "mmluPro": 55.8, "eqBench": 77.4, "ifeval": 76.4, "hfDownloads30d": 5028, "hfLikes": 76, "ollamaCommand": "dphn/dolphin-2.9.4-llama3.1-8b-gguf", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-unslopnemo-12b-v4-1-gguf", "name": "UnslopNemo-12B-v4.1-GGUF", "hfRepo": "TheDrummer/UnslopNemo-12B-v4.1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-10-23", "lastVerified": "2026-11-01", "ugiScore": 67.8, "complianceRate": 98.7, "mmluPro": 57.2, "eqBench": 82.1, "ifeval": 78.1, "hfDownloads30d": 5447, "hfLikes": 36, "ollamaCommand": "TheDrummer/UnslopNemo-12B-v4.1-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-12b-it-abliterated", "name": "Huihui-gemma-4-12B-it-abliterated", "hfRepo": "huihui-ai/Huihui-gemma-4-12B-it-abliterated", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-06", "lastVerified": "2026-06-07", "ugiScore": 68.1, "complianceRate": 97.8, "mmluPro": 58.8, "eqBench": 77.7, "ifeval": 78.4, "hfDownloads30d": 2420, "hfLikes": 228, "ollamaCommand": "huihui-ai/Huihui-gemma-4-12B-it-abliterated", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 2 (12B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nous-hermes-3-llama-3.1-405b", "name": "Hermes-3-Llama-3.1-405B", "hfRepo": "NousResearch/Hermes-3-Llama-3.1-405B", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (Dense 405B)", "paramsBillions": 405, "parameterClass": "Frontier (100B+)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-19", "lastVerified": "2024-10-08", "ugiScore": 78.2, "complianceRate": 96.8, "mmluPro": 76.5, "eqBench": 88.7, "ifeval": 89, "hfDownloads30d": 1654, "hfLikes": 276, "openRouterSlug": "nousresearch/hermes-3-llama-3.1-405b", "ollamaCommand": "NousResearch/Hermes-3-Llama-3.1-405B", "vramQ4GB": 228, "vramQ8GB": 415, "vramFP16GB": 810, "summary": "Flagship frontier open-weight model fine-tuned by Nous Research for neutral user alignment, complex agentic tool calling, and deep creative synthesis.", "strengths": [ "Highest overall UGI (78.2) & MMLU-Pro (76.5)", "Native function calling & JSON schema adherence" ], "tradeoffs": [ "Cloud API or multi-node H100 cluster required (228GB VRAM at Q4)" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-4-70b", "name": "Hermes-4-70B", "hfRepo": "NousResearch/Hermes-4-70B", "creator": "NousResearch", "baseArchitecture": "Open Transformer (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-08-18", "lastVerified": "2025-09-02", "ugiScore": 77.4, "complianceRate": 99.6, "mmluPro": 71, "eqBench": 86.8, "ifeval": 87, "hfDownloads30d": 2354, "hfLikes": 227, "ollamaCommand": "NousResearch/Hermes-4-70B", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (70B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (40 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "thedrummer-gemmasutra-small-4b-v1-gguf", "name": "Gemmasutra-Small-4B-v1-GGUF", "hfRepo": "TheDrummer/Gemmasutra-Small-4B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2025-03-11", "lastVerified": "2026-11-01", "ugiScore": 61.7, "complianceRate": 99.6, "mmluPro": 50.1, "eqBench": 76.9, "ifeval": 71.9, "hfDownloads30d": 5507, "hfLikes": 15, "ollamaCommand": "TheDrummer/Gemmasutra-Small-4B-v1-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (4B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~3.4 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "thedrummer-unslopnemo-12b-v3-gguf", "name": "UnslopNemo-12B-v3-GGUF", "hfRepo": "TheDrummer/UnslopNemo-12B-v3-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-10-10", "lastVerified": "2026-11-01", "ugiScore": 66.6, "complianceRate": 98.9, "mmluPro": 57.6, "eqBench": 81.3, "ifeval": 76.9, "hfDownloads30d": 5053, "hfLikes": 38, "ollamaCommand": "TheDrummer/UnslopNemo-12B-v3-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-theta-llama-3-8b-gguf", "name": "Hermes-2-Theta-Llama-3-8B-GGUF", "hfRepo": "NousResearch/Hermes-2-Theta-Llama-3-8B-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-13", "lastVerified": "2024-05-14", "ugiScore": 66.5, "complianceRate": 98.6, "mmluPro": 55.5, "eqBench": 77.7, "ifeval": 75.8, "hfDownloads30d": 4211, "hfLikes": 93, "ollamaCommand": "NousResearch/Hermes-2-Theta-Llama-3-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-6-27b-abliterated", "name": "Huihui-Qwen3.6-27B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.6-27B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-23", "lastVerified": "2026-04-23", "ugiScore": 71.6, "complianceRate": 99, "mmluPro": 67.8, "eqBench": 82.3, "ifeval": 82.5, "hfDownloads30d": 4413, "hfLikes": 75, "ollamaCommand": "huihui-ai/Huihui-Qwen3.6-27B-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.6 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "orenguteng-llama-3-8b-lexi-uncensored", "name": "Llama-3-8B-Lexi-Uncensored", "hfRepo": "Orenguteng/Llama-3-8B-Lexi-Uncensored", "creator": "Orenguteng", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-04-23", "lastVerified": "2026-11-01", "ugiScore": 65.6, "complianceRate": 99.3, "mmluPro": 56, "eqBench": 75.8, "ifeval": 74.9, "hfDownloads30d": 951, "hfLikes": 298, "ollamaCommand": "Orenguteng/Llama-3-8B-Lexi-Uncensored", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by Orenguteng, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "jartine-dolphin-2-5-mixtral-8x7b-llamafile", "name": "dolphin-2.5-mixtral-8x7b-llamafile", "hfRepo": "jartine/dolphin-2.5-mixtral-8x7b-llamafile", "creator": "jartine", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-12-28", "lastVerified": "2026-11-01", "ugiScore": 73.7, "complianceRate": 98.8, "mmluPro": 67.7, "eqBench": 84.4, "ifeval": 83.3, "hfDownloads30d": 5345, "hfLikes": 4, "ollamaCommand": "jartine/dolphin-2.5-mixtral-8x7b-llamafile", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by jartine, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v4-1", "name": "Cydonia-24B-v4.1", "hfRepo": "TheDrummer/Cydonia-24B-v4.1", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-08-17", "lastVerified": "2025-08-24", "ugiScore": 70.6, "complianceRate": 99.7, "mmluPro": 62.5, "eqBench": 84.8, "ifeval": 81, "hfDownloads30d": 4104, "hfLikes": 83, "ollamaCommand": "TheDrummer/Cydonia-24B-v4.1", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-4b-instruct-2507-abliterated", "name": "Huihui-Qwen3-4B-Instruct-2507-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-4B-Instruct-2507-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-08-07", "lastVerified": "2026-11-01", "ugiScore": 64.4, "complianceRate": 99.3, "mmluPro": 56, "eqBench": 73.1, "ifeval": 75.1, "hfDownloads30d": 4657, "hfLikes": 45, "ollamaCommand": "huihui-ai/Huihui-Qwen3-4B-Instruct-2507-abliterated", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v4-3", "name": "Cydonia-24B-v4.3", "hfRepo": "TheDrummer/Cydonia-24B-v4.3", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-11-08", "lastVerified": "2025-12-17", "ugiScore": 70.8, "complianceRate": 98.7, "mmluPro": 62.2, "eqBench": 85.3, "ifeval": 81.2, "hfDownloads30d": 2491, "hfLikes": 187, "ollamaCommand": "TheDrummer/Cydonia-24B-v4.3", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "isvalorum-huihui-qwen3-8-27b-coder390-efficientthink-abliterated-val-apex-i-miniplus-v3-gguf", "name": "Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-VAL-APEX-I-MiniPlus-V3-GGUF", "hfRepo": "IsValorum/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-VAL-APEX-I-MiniPlus-V3-GGUF", "creator": "IsValorum", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-14", "lastVerified": "2026-11-01", "ugiScore": 72.4, "complianceRate": 99.2, "mmluPro": 70.3, "eqBench": 80.2, "ifeval": 85.5, "hfDownloads30d": 5127, "hfLikes": 5, "ollamaCommand": "IsValorum/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-VAL-APEX-I-MiniPlus-V3-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by IsValorum, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nyan-ame-qwen2-5-coder-7b-instruct-abliterated-gguf", "name": "Qwen2.5-Coder-7B-Instruct-Abliterated-GGUF", "hfRepo": "nyan-ame/Qwen2.5-Coder-7B-Instruct-Abliterated-GGUF", "creator": "nyan-ame", "baseArchitecture": "Qwen 2.5 Coder (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-17", "lastVerified": "2026-11-01", "ugiScore": 66.8, "complianceRate": 97.9, "mmluPro": 61.6, "eqBench": 73, "ifeval": 79.7, "hfDownloads30d": 5125, "hfLikes": 1, "ollamaCommand": "nyan-ame/Qwen2.5-Coder-7B-Instruct-Abliterated-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (7B) checkpoint by nyan-ame, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davidau-l3-stheno-maid-blackroot-grand-horror-16b-gguf", "name": "L3-Stheno-Maid-Blackroot-Grand-HORROR-16B-GGUF", "hfRepo": "DavidAU/L3-Stheno-Maid-Blackroot-Grand-HORROR-16B-GGUF", "creator": "DavidAU", "baseArchitecture": "Llama 3.1 (16B)", "paramsBillions": 16, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-07-03", "lastVerified": "2026-11-01", "ugiScore": 69.7, "complianceRate": 99.3, "mmluPro": 59.1, "eqBench": 83.4, "ifeval": 80.1, "hfDownloads30d": 3457, "hfLikes": 109, "ollamaCommand": "DavidAU/L3-Stheno-Maid-Blackroot-Grand-HORROR-16B-GGUF", "vramQ4GB": 10, "vramQ8GB": 19, "vramFP16GB": 35, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (16B) by DavidAU, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~10 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "davidau-tiefighter-holodeck-holomax-mythomax-f1-v1-compos-20b-gguf", "name": "TieFighter-Holodeck-Holomax-Mythomax-F1-V1-COMPOS-20B-gguf", "hfRepo": "DavidAU/TieFighter-Holodeck-Holomax-Mythomax-F1-V1-COMPOS-20B-gguf", "creator": "DavidAU", "baseArchitecture": "Open Transformer (20B)", "paramsBillions": 20, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-05-18", "lastVerified": "2026-11-01", "ugiScore": 70.2, "complianceRate": 99.7, "mmluPro": 60.7, "eqBench": 84.3, "ifeval": 80.5, "hfDownloads30d": 4047, "hfLikes": 68, "ollamaCommand": "DavidAU/TieFighter-Holodeck-Holomax-Mythomax-F1-V1-COMPOS-20B-gguf", "vramQ4GB": 12, "vramQ8GB": 23, "vramFP16GB": 44, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (20B) by DavidAU, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-1-7b-abliterated-v2", "name": "Huihui-Qwen3-1.7B-abliterated-v2", "hfRepo": "huihui-ai/Huihui-Qwen3-1.7B-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (1.7B)", "paramsBillions": 1.7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-19", "lastVerified": "2026-11-01", "ugiScore": 58.8, "complianceRate": 98.8, "mmluPro": 48.7, "eqBench": 70.4, "ifeval": 69.4, "hfDownloads30d": 4885, "hfLikes": 6, "ollamaCommand": "huihui-ai/Huihui-Qwen3-1.7B-abliterated-v2", "vramQ4GB": 2.2, "vramQ8GB": 3.1, "vramFP16GB": 5, "summary": "Orthogonal weight-projected Qwen 3 (1.7B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (1.7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1.7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-mistral-small-3-2-24b-instruct-2506-abliterated", "name": "Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated", "hfRepo": "huihui-ai/Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated", "creator": "huihui-ai", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-07-02", "lastVerified": "2026-11-01", "ugiScore": 71.3, "complianceRate": 99, "mmluPro": 63.5, "eqBench": 81.5, "ifeval": 81.7, "hfDownloads30d": 4594, "hfLikes": 25, "ollamaCommand": "huihui-ai/Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Orthogonal weight-projected Mistral Small 3 (24B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB VRAM at Q4_K_M)", "Preserves native Mistral Small 3 (24B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 24B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "deepseek-r1-distill-qwen-32b-uncensored", "name": "DeepSeek-R1-Distill-Qwen-32B-abliterated", "hfRepo": "huihui-ai/DeepSeek-R1-Distill-Qwen-32B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 32B (R1 Distill)", "paramsBillions": 32.8, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-29", "lastVerified": "2026-11-01", "ugiScore": 74.9, "complianceRate": 97.4, "mmluPro": 74.2, "eqBench": 81.8, "ifeval": 83.9, "hfDownloads30d": 1061, "hfLikes": 260, "openRouterSlug": "deepseek/deepseek-r1-distill-qwen-32b", "ollamaCommand": "huihui-ai/DeepSeek-R1-Distill-Qwen-32B-abliterated", "vramQ4GB": 20, "vramQ8GB": 35, "vramFP16GB": 66, "summary": "Removes both alignment refusals and political/topic guardrails from DeepSeek R1 32B Distill, delivering frontier-grade reasoning on a single 24GB GPU.", "strengths": [ "74.2 MMLU-Pro reasoning score on 24GB VRAM", "MIT License", "Unfiltered chain-of-thought ()" ], "tradeoffs": [ "Verbose thinking tokens increase time-to-first-answer" ], "isLiveSynced": true }, { "id": "mlabonne-hermes-3-llama-3-1-8b-lorablated-gguf", "name": "Hermes-3-Llama-3.1-8B-lorablated-GGUF", "hfRepo": "mlabonne/Hermes-3-Llama-3.1-8B-lorablated-GGUF", "creator": "mlabonne", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-16", "lastVerified": "2024-08-16", "ugiScore": 68.6, "complianceRate": 98.4, "mmluPro": 56.5, "eqBench": 78.1, "ifeval": 79.2, "hfDownloads30d": 4206, "hfLikes": 45, "ollamaCommand": "mlabonne/Hermes-3-Llama-3.1-8B-lorablated-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by mlabonne, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "neversleep-lumimaid-v0-2-12b-gguf", "name": "Lumimaid-v0.2-12B-GGUF", "hfRepo": "NeverSleep/Lumimaid-v0.2-12B-GGUF", "creator": "NeverSleep", "baseArchitecture": "Llama 3.1 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-07-27", "lastVerified": "2026-11-01", "ugiScore": 66.1, "complianceRate": 99.6, "mmluPro": 58.1, "eqBench": 82.3, "ifeval": 76.4, "hfDownloads30d": 4189, "hfLikes": 44, "ollamaCommand": "NeverSleep/Lumimaid-v0.2-12B-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (12B) by NeverSleep, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "insraq-qwen3-5-4b-emperoai-qwen3-8-distill-heretic-abliterated-mtp-gguf", "name": "Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated-MTP-GGUF", "hfRepo": "insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated-MTP-GGUF", "creator": "insraq", "baseArchitecture": "Qwen 3.8 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-18", "lastVerified": "2026-11-01", "ugiScore": 62.8, "complianceRate": 98.3, "mmluPro": 55.4, "eqBench": 74.7, "ifeval": 73.5, "hfDownloads30d": 4449, "hfLikes": 15, "ollamaCommand": "insraq/Qwen3.5-4B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated-MTP-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3.8 (4B) checkpoint by insraq, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "vcruz305-muse-glimmer-30b-hermes-agentic-gguf", "name": "Muse-Glimmer-30B-Hermes-Agentic-GGUF", "hfRepo": "vcruz305/Muse-Glimmer-30B-Hermes-Agentic-GGUF", "creator": "vcruz305", "baseArchitecture": "Open Transformer (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-14", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 99.4, "mmluPro": 63.8, "eqBench": 82.5, "ifeval": 82, "hfDownloads30d": 4591, "hfLikes": 1, "ollamaCommand": "vcruz305/Muse-Glimmer-30B-Hermes-Agentic-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (30B) by vcruz305, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~18 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 30B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-skyfall-36b-v2", "name": "Skyfall-36B-v2", "hfRepo": "TheDrummer/Skyfall-36B-v2", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (36B)", "paramsBillions": 36, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-02-03", "lastVerified": "2026-11-01", "ugiScore": 72.1, "complianceRate": 98.4, "mmluPro": 65.6, "eqBench": 86.5, "ifeval": 82.6, "hfDownloads30d": 3781, "hfLikes": 54, "ollamaCommand": "TheDrummer/Skyfall-36B-v2", "vramQ4GB": 21, "vramQ8GB": 40, "vramFP16GB": 78, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (36B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-6-35b-a3b-abliterated", "name": "Huihui-Qwen3.6-35B-A3B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-18", "lastVerified": "2026-04-18", "ugiScore": 73.8, "complianceRate": 98.8, "mmluPro": 69.3, "eqBench": 82.5, "ifeval": 84.7, "hfDownloads30d": 2973, "hfLikes": 104, "ollamaCommand": "huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gpt-oss-20b-bf16-abliterated-v2", "name": "Huihui-gpt-oss-20b-BF16-abliterated-v2", "hfRepo": "huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (20B)", "paramsBillions": 20, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-09-27", "lastVerified": "2026-11-01", "ugiScore": 68.8, "complianceRate": 98.5, "mmluPro": 62.1, "eqBench": 80.7, "ifeval": 79.1, "hfDownloads30d": 3773, "hfLikes": 44, "ollamaCommand": "huihui-ai/Huihui-gpt-oss-20b-BF16-abliterated-v2", "vramQ4GB": 12, "vramQ8GB": 23, "vramFP16GB": 44, "summary": "Orthogonal weight-projected Open Transformer (20B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (20B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 20B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-6-35b-a3b-claude-4-7-opus-abliterated", "name": "Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-21", "lastVerified": "2026-04-21", "ugiScore": 73.3, "complianceRate": 98.3, "mmluPro": 69.4, "eqBench": 82.1, "ifeval": 84.2, "hfDownloads30d": 1016, "hfLikes": 223, "ollamaCommand": "huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.6 (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-3-llama-3-1-70b", "name": "Hermes-3-Llama-3.1-70B", "hfRepo": "NousResearch/Hermes-3-Llama-3.1-70B", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-07-29", "lastVerified": "2024-09-08", "ugiScore": 78.3, "complianceRate": 99, "mmluPro": 69.9, "eqBench": 87.6, "ifeval": 89.2, "hfDownloads30d": 2175, "hfLikes": 145, "ollamaCommand": "NousResearch/Hermes-3-Llama-3.1-70B", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (70B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (40 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "dphn-dolphin-2-9-2-qwen2-7b-gguf", "name": "dolphin-2.9.2-qwen2-7b-gguf", "hfRepo": "dphn/dolphin-2.9.2-qwen2-7b-gguf", "creator": "dphn", "baseArchitecture": "Qwen 2.5 (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-24", "lastVerified": "2026-11-01", "ugiScore": 65.8, "complianceRate": 98.5, "mmluPro": 54.2, "eqBench": 75.4, "ifeval": 75.1, "hfDownloads30d": 3615, "hfLikes": 47, "ollamaCommand": "dphn/dolphin-2.9.2-qwen2-7b-gguf", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Qwen 2.5 (7B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "sao10k-l3-8b-lunaris-v1", "name": "L3-8B-Lunaris-v1", "hfRepo": "Sao10K/L3-8B-Lunaris-v1", "creator": "Sao10K", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-06-26", "lastVerified": "2024-06-29", "ugiScore": 66.1, "complianceRate": 98.8, "mmluPro": 54.9, "eqBench": 78.9, "ifeval": 76.3, "hfDownloads30d": 2067, "hfLikes": 147, "ollamaCommand": "Sao10K/L3-8B-Lunaris-v1", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (8B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-mistral-7b-dpo-gguf", "name": "Nous-Hermes-2-Mistral-7B-DPO-GGUF", "hfRepo": "NousResearch/Nous-Hermes-2-Mistral-7B-DPO-GGUF", "creator": "NousResearch", "baseArchitecture": "Mistral (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-02-20", "lastVerified": "2024-02-21", "ugiScore": 66.4, "complianceRate": 98.3, "mmluPro": 54.9, "eqBench": 76.2, "ifeval": 75.7, "hfDownloads30d": 2784, "hfLikes": 95, "ollamaCommand": "NousResearch/Nous-Hermes-2-Mistral-7B-DPO-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Mistral (7B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-122b-a10b-abliterated-gguf", "name": "Huihui-Qwen3.5-122B-A10B-abliterated-GGUF", "hfRepo": "huihui-ai/Huihui-Qwen3.5-122B-A10B-abliterated-GGUF", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (122B)", "paramsBillions": 122, "parameterClass": "Frontier (100B+)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-11", "lastVerified": "2026-03-11", "ugiScore": 75.9, "complianceRate": 98.2, "mmluPro": 74.7, "eqBench": 86.4, "ifeval": 86.9, "hfDownloads30d": 3711, "hfLikes": 33, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-122B-A10B-abliterated-GGUF", "vramQ4GB": 69, "vramQ8GB": 134, "vramFP16GB": 262, "summary": "Orthogonal weight-projected Qwen 3.5 (122B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on multi-GPU workstations or cloud nodes.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for multi-GPU workstations or cloud nodes (~69 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (122B) instruction and coding schemas" ], "tradeoffs": [ "Requires 69 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "dphn-dolphin-x1-8b-gguf", "name": "Dolphin-X1-8B-GGUF", "hfRepo": "dphn/Dolphin-X1-8B-GGUF", "creator": "dphn", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-09-18", "lastVerified": "2026-11-01", "ugiScore": 67.3, "complianceRate": 99.4, "mmluPro": 56, "eqBench": 76.6, "ifeval": 76.6, "hfDownloads30d": 3710, "hfLikes": 33, "ollamaCommand": "dphn/Dolphin-X1-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (8B) by dphn, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-artemis-31b-v1-1-gguf", "name": "Artemis-31B-v1.1-GGUF", "hfRepo": "TheDrummer/Artemis-31B-v1.1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-07-13", "lastVerified": "2026-08-06", "ugiScore": 71.7, "complianceRate": 99.6, "mmluPro": 63.7, "eqBench": 85.4, "ifeval": 82.1, "hfDownloads30d": 3614, "hfLikes": 33, "ollamaCommand": "TheDrummer/Artemis-31B-v1.1-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (31B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-mistral-7b-dpo", "name": "Nous-Hermes-2-Mistral-7B-DPO", "hfRepo": "NousResearch/Nous-Hermes-2-Mistral-7B-DPO", "creator": "NousResearch", "baseArchitecture": "Mistral (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-02-18", "lastVerified": "2024-04-30", "ugiScore": 65.8, "complianceRate": 99.3, "mmluPro": 55.4, "eqBench": 75.7, "ifeval": 75.1, "hfDownloads30d": 752, "hfLikes": 222, "ollamaCommand": "NousResearch/Nous-Hermes-2-Mistral-7B-DPO", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Mistral (7B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "heyujie-qwen3-5-27b-abliterated-gguf", "name": "Qwen3.5-27B-abliterated-GGUF", "hfRepo": "HeYujie/Qwen3.5-27B-abliterated-GGUF", "creator": "HeYujie", "baseArchitecture": "Qwen 3.5 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-02-28", "lastVerified": "2026-11-01", "ugiScore": 71.8, "complianceRate": 99.2, "mmluPro": 68.7, "eqBench": 82.3, "ifeval": 82.7, "hfDownloads30d": 4004, "hfLikes": 1, "ollamaCommand": "HeYujie/Qwen3.5-27B-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.5 (27B) checkpoint by HeYujie, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sevenofnine-gemma-4-26b-a4b-it-abliterated-gguf", "name": "Gemma-4-26B-A4B-It-Abliterated-GGUF", "hfRepo": "SevenOfNine/Gemma-4-26B-A4B-It-Abliterated-GGUF", "creator": "SevenOfNine", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-10", "lastVerified": "2026-11-01", "ugiScore": 71.6, "complianceRate": 98.9, "mmluPro": 64.8, "eqBench": 82.1, "ifeval": 81.9, "hfDownloads30d": 3933, "hfLikes": 5, "ollamaCommand": "SevenOfNine/Gemma-4-26B-A4B-It-Abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by SevenOfNine, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "melvin56-deepseek-r1-distill-qwen-7b-abliterated-v2-gguf", "name": "DeepSeek-R1-Distill-Qwen-7B-abliterated-v2-GGUF", "hfRepo": "Melvin56/DeepSeek-R1-Distill-Qwen-7B-abliterated-v2-GGUF", "creator": "Melvin56", "baseArchitecture": "DeepSeek R1 Distill Qwen (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-31", "lastVerified": "2026-11-01", "ugiScore": 65.6, "complianceRate": 99.1, "mmluPro": 58.7, "eqBench": 75.3, "ifeval": 76.3, "hfDownloads30d": 3667, "hfLikes": 13, "ollamaCommand": "Melvin56/DeepSeek-R1-Distill-Qwen-7B-abliterated-v2-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (7B) checkpoint by Melvin56, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-pro-llama-3-8b-gguf", "name": "Hermes-2-Pro-Llama-3-8B-GGUF", "hfRepo": "NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-01", "lastVerified": "2024-05-03", "ugiScore": 66.2, "complianceRate": 99.1, "mmluPro": 54.9, "eqBench": 77.3, "ifeval": 75.5, "hfDownloads30d": 1365, "hfLikes": 165, "ollamaCommand": "NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "isvalorum-huihui-qwen3-8-27b-coder390-efficientthink-abliterated-val-apex-i-nanoplus-v3-gguf", "name": "Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-VAL-APEX-I-NanoPlus-V3-GGUF", "hfRepo": "IsValorum/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-VAL-APEX-I-NanoPlus-V3-GGUF", "creator": "IsValorum", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-14", "lastVerified": "2026-11-01", "ugiScore": 72.3, "complianceRate": 99.1, "mmluPro": 69.5, "eqBench": 80.1, "ifeval": 85.4, "hfDownloads30d": 3758, "hfLikes": 4, "ollamaCommand": "IsValorum/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-VAL-APEX-I-NanoPlus-V3-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by IsValorum, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-fimbulvetr-11b-v2-gguf", "name": "Fimbulvetr-11B-v2-GGUF", "hfRepo": "Sao10K/Fimbulvetr-11B-v2-GGUF", "creator": "Sao10K", "baseArchitecture": "Open Transformer (11B)", "paramsBillions": 11, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-02-06", "lastVerified": "2024-03-02", "ugiScore": 66.8, "complianceRate": 98.7, "mmluPro": 56.4, "eqBench": 80.4, "ifeval": 77.1, "hfDownloads30d": 2000, "hfLikes": 121, "ollamaCommand": "Sao10K/Fimbulvetr-11B-v2-GGUF", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 25, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (11B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "arliai-mistral-nemo-12b-arliai-rpmax-v1-3-gguf", "name": "Mistral-Nemo-12B-ArliAI-RPMax-v1.3-GGUF", "hfRepo": "ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.3-GGUF", "creator": "ArliAI", "baseArchitecture": "Mistral NeMo (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-12-05", "lastVerified": "2026-11-01", "ugiScore": 67.7, "complianceRate": 98.8, "mmluPro": 59.2, "eqBench": 77.6, "ifeval": 77.1, "hfDownloads30d": 3471, "hfLikes": 16, "ollamaCommand": "ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.3-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Mistral NeMo (12B) by ArliAI, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-llama-3-2-3b-instruct-abliterated", "name": "Llama-3.2-3B-Instruct-abliterated", "hfRepo": "huihui-ai/Llama-3.2-3B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Llama 3.2 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-09-28", "lastVerified": "2026-11-01", "ugiScore": 61.1, "complianceRate": 98.9, "mmluPro": 49.3, "eqBench": 72.4, "ifeval": 71.2, "hfDownloads30d": 2089, "hfLikes": 104, "ollamaCommand": "huihui-ai/Llama-3.2-3B-Instruct-abliterated", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Orthogonal weight-projected Llama 3.2 (3B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.9 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (3B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 3B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "kepom-qwen2-5-7b-instruct-abliterated-v2-gguf", "name": "Qwen2.5-7B-Instruct-abliterated-v2-GGUF", "hfRepo": "kepom/Qwen2.5-7B-Instruct-abliterated-v2-GGUF", "creator": "kepom", "baseArchitecture": "Qwen 2.5 (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-14", "lastVerified": "2026-11-01", "ugiScore": 65.3, "complianceRate": 98.2, "mmluPro": 59.4, "eqBench": 76.9, "ifeval": 76, "hfDownloads30d": 3603, "hfLikes": 1, "ollamaCommand": "kepom/Qwen2.5-7B-Instruct-abliterated-v2-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 (7B) checkpoint by kepom, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "crusoeai-dolphin-2-9-3-qwen2-1-5b-gguf", "name": "dolphin-2.9.3-qwen2-1.5b-GGUF", "hfRepo": "crusoeai/dolphin-2.9.3-qwen2-1.5b-GGUF", "creator": "crusoeai", "baseArchitecture": "Qwen 2.5 (1.5B)", "paramsBillions": 1.5, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-06-11", "lastVerified": "2026-11-01", "ugiScore": 59, "complianceRate": 99.6, "mmluPro": 45.6, "eqBench": 69.3, "ifeval": 68.2, "hfDownloads30d": 3572, "hfLikes": 3, "ollamaCommand": "crusoeai/dolphin-2.9.3-qwen2-1.5b-GGUF", "vramQ4GB": 2.1, "vramQ8GB": 2.9, "vramFP16GB": 5, "summary": "Uncensored DPO/SFT fine-tune of Qwen 2.5 (1.5B) by crusoeai, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 1.5B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "arliai-mistral-nemo-12b-arliai-rpmax-v1-2-gguf", "name": "Mistral-Nemo-12B-ArliAI-RPMax-v1.2-GGUF", "hfRepo": "ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2-GGUF", "creator": "ArliAI", "baseArchitecture": "Mistral NeMo (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-10-12", "lastVerified": "2026-11-01", "ugiScore": 67, "complianceRate": 98.9, "mmluPro": 58.3, "eqBench": 77.5, "ifeval": 76.4, "hfDownloads30d": 3205, "hfLikes": 23, "ollamaCommand": "ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Mistral NeMo (12B) by ArliAI, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-4-405b", "name": "Hermes-4-405B", "hfRepo": "NousResearch/Hermes-4-405B", "creator": "NousResearch", "baseArchitecture": "Open Transformer (405B)", "paramsBillions": 405, "parameterClass": "Frontier (100B+)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-08-06", "lastVerified": "2025-09-02", "ugiScore": 77.3, "complianceRate": 98.6, "mmluPro": 71, "eqBench": 85.2, "ifeval": 86.9, "hfDownloads30d": 2065, "hfLikes": 97, "ollamaCommand": "NousResearch/Hermes-4-405B", "vramQ4GB": 228, "vramQ8GB": 440, "vramFP16GB": 866, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (405B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on multi-GPU workstations or cloud nodes (~228 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (228 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "zetasepic-qwen2-5-32b-instruct-abliterated-v2-gguf", "name": "Qwen2.5-32B-Instruct-abliterated-v2-GGUF", "hfRepo": "zetasepic/Qwen2.5-32B-Instruct-abliterated-v2-GGUF", "creator": "zetasepic", "baseArchitecture": "Qwen 2.5 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-10-11", "lastVerified": "2026-11-01", "ugiScore": 72.6, "complianceRate": 99.2, "mmluPro": 69, "eqBench": 83.2, "ifeval": 83.5, "hfDownloads30d": 3169, "hfLikes": 21, "ollamaCommand": "zetasepic/Qwen2.5-32B-Instruct-abliterated-v2-GGUF", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 2.5 (32B) checkpoint by zetasepic, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "luffythefox-tiel-coder-35b-a3b-genesis-hermes-gguf", "name": "Tiel-Coder-35B-A3B-Genesis-Hermes-GGUF", "hfRepo": "LuffyTheFox/Tiel-Coder-35B-A3B-Genesis-Hermes-GGUF", "creator": "LuffyTheFox", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-08-29", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.7, "mmluPro": 69.1, "eqBench": 80.1, "ifeval": 84.2, "hfDownloads30d": 3021, "hfLikes": 29, "ollamaCommand": "LuffyTheFox/Tiel-Coder-35B-A3B-Genesis-Hermes-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (35B) by LuffyTheFox, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-mixtral-8x7b-dpo-gguf", "name": "Nous-Hermes-2-Mixtral-8x7B-DPO-GGUF", "hfRepo": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO-GGUF", "creator": "NousResearch", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-16", "lastVerified": "2026-11-01", "ugiScore": 74.5, "complianceRate": 99.6, "mmluPro": 66.9, "eqBench": 84.6, "ifeval": 84.1, "hfDownloads30d": 2316, "hfLikes": 76, "ollamaCommand": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO-GGUF", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-4b-it-abliterated-v2-gguf", "name": "gemma-3-4b-it-abliterated-v2-GGUF", "hfRepo": "mlabonne/gemma-3-4b-it-abliterated-v2-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-05-28", "lastVerified": "2025-05-29", "ugiScore": 62.2, "complianceRate": 98.1, "mmluPro": 51.5, "eqBench": 75.5, "ifeval": 72.4, "hfDownloads30d": 3225, "hfLikes": 15, "ollamaCommand": "mlabonne/gemma-3-4b-it-abliterated-v2-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Gemma 3 (4B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-12b-it-abliterated-gguf", "name": "gemma-3-12b-it-abliterated-GGUF", "hfRepo": "mlabonne/gemma-3-12b-it-abliterated-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-03-17", "lastVerified": "2025-03-18", "ugiScore": 66.4, "complianceRate": 99.1, "mmluPro": 59.6, "eqBench": 81.1, "ifeval": 76.7, "hfDownloads30d": 2567, "hfLikes": 55, "ollamaCommand": "mlabonne/gemma-3-12b-it-abliterated-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 3 (12B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "brittlewis12-kunoichi-dpo-v2-7b-gguf", "name": "Kunoichi-DPO-v2-7B-GGUF", "hfRepo": "brittlewis12/Kunoichi-DPO-v2-7B-GGUF", "creator": "brittlewis12", "baseArchitecture": "Open Transformer (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-16", "lastVerified": "2026-11-01", "ugiScore": 66.7, "complianceRate": 99.4, "mmluPro": 54.7, "eqBench": 76.6, "ifeval": 76, "hfDownloads30d": 2053, "hfLikes": 89, "ollamaCommand": "brittlewis12/Kunoichi-DPO-v2-7B-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (7B) by brittlewis12, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-4b-abliterated-v2", "name": "Huihui-Qwen3-4B-abliterated-v2", "hfRepo": "huihui-ai/Huihui-Qwen3-4B-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-19", "lastVerified": "2026-11-01", "ugiScore": 63.5, "complianceRate": 97.8, "mmluPro": 55.9, "eqBench": 73.5, "ifeval": 74.2, "hfDownloads30d": 2396, "hfLikes": 66, "ollamaCommand": "huihui-ai/Huihui-Qwen3-4B-abliterated-v2", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-14b-abliterated-v2", "name": "Huihui-Qwen3-14B-abliterated-v2", "hfRepo": "huihui-ai/Huihui-Qwen3-14B-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-17", "lastVerified": "2026-11-01", "ugiScore": 70.3, "complianceRate": 98.9, "mmluPro": 64, "eqBench": 79.6, "ifeval": 81.1, "hfDownloads30d": 2799, "hfLikes": 38, "ollamaCommand": "huihui-ai/Huihui-Qwen3-14B-abliterated-v2", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected Qwen 3 (14B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-deepseek-r1-distill-qwen-14b-abliterated-v2", "name": "DeepSeek-R1-Distill-Qwen-14B-abliterated-v2", "hfRepo": "huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "DeepSeek R1 Distill Qwen (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-23", "lastVerified": "2026-11-01", "ugiScore": 68.6, "complianceRate": 97.8, "mmluPro": 64.4, "eqBench": 78.2, "ifeval": 79.4, "hfDownloads30d": 824, "hfLikes": 168, "ollamaCommand": "huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-v2", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (14B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "jan1k-qwen3-6-35b-a3b-uncensored-genesis-hermes-final-nvfp4-gguf", "name": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-NVFP4-GGUF", "hfRepo": "jan1k/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-NVFP4-GGUF", "creator": "jan1k", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-09-12", "lastVerified": "2026-11-01", "ugiScore": 75.1, "complianceRate": 98.1, "mmluPro": 69.1, "eqBench": 82.2, "ifeval": 85.1, "hfDownloads30d": 3159, "hfLikes": 12, "ollamaCommand": "jan1k/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-NVFP4-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.6 (35B) by jan1k, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-rocinante-x-12b-v1", "name": "Rocinante-X-12B-v1", "hfRepo": "TheDrummer/Rocinante-X-12B-v1", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-01-19", "lastVerified": "2026-01-25", "ugiScore": 66.8, "complianceRate": 98.9, "mmluPro": 57.2, "eqBench": 81, "ifeval": 77.1, "hfDownloads30d": 1592, "hfLikes": 108, "ollamaCommand": "TheDrummer/Rocinante-X-12B-v1", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-4b-it-abliterated-gguf", "name": "gemma-3-4b-it-abliterated-GGUF", "hfRepo": "mlabonne/gemma-3-4b-it-abliterated-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-03-17", "lastVerified": "2025-03-17", "ugiScore": 61.1, "complianceRate": 97.6, "mmluPro": 52.1, "eqBench": 76.3, "ifeval": 71.3, "hfDownloads30d": 2650, "hfLikes": 36, "ollamaCommand": "mlabonne/gemma-3-4b-it-abliterated-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Gemma 3 (4B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-tiger-gemma-12b-v3-gguf", "name": "Tiger-Gemma-12B-v3-GGUF", "hfRepo": "TheDrummer/Tiger-Gemma-12B-v3-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2025-06-28", "lastVerified": "2026-11-01", "ugiScore": 68.1, "complianceRate": 99.6, "mmluPro": 58.1, "eqBench": 80.9, "ifeval": 78.4, "hfDownloads30d": 2824, "hfLikes": 23, "ollamaCommand": "TheDrummer/Tiger-Gemma-12B-v3-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "sao10k-fimbulvetr-11b-v2", "name": "Fimbulvetr-11B-v2", "hfRepo": "Sao10K/Fimbulvetr-11B-v2", "creator": "Sao10K", "baseArchitecture": "Open Transformer (11B)", "paramsBillions": 11, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-02-06", "lastVerified": "2024-04-04", "ugiScore": 67.8, "complianceRate": 99.5, "mmluPro": 56.5, "eqBench": 81.6, "ifeval": 78.1, "hfDownloads30d": 230, "hfLikes": 195, "ollamaCommand": "Sao10K/Fimbulvetr-11B-v2", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 25, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (11B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "thedrummer-rocinante-12b-v1-1", "name": "Rocinante-12B-v1.1", "hfRepo": "TheDrummer/Rocinante-12B-v1.1", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-08-15", "lastVerified": "2026-11-01", "ugiScore": 66.4, "complianceRate": 99.5, "mmluPro": 56.8, "eqBench": 82.1, "ifeval": 76.7, "hfDownloads30d": 1151, "hfLikes": 131, "ollamaCommand": "TheDrummer/Rocinante-12B-v1.1", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "interimlabs-interimlabs-huihui-qwen3-5-9b-abliterated-gguf-q4-k-m", "name": "InterimLabs-Huihui-Qwen3.5-9B-abliterated-GGUF-Q4_K_M", "hfRepo": "interimlabs/InterimLabs-Huihui-Qwen3.5-9B-abliterated-GGUF-Q4_K_M", "creator": "interimlabs", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-01", "lastVerified": "2026-11-01", "ugiScore": 67.3, "complianceRate": 99.1, "mmluPro": 60, "eqBench": 78.1, "ifeval": 78.1, "hfDownloads30d": 3092, "hfLikes": 1, "ollamaCommand": "interimlabs/InterimLabs-Huihui-Qwen3.5-9B-abliterated-GGUF-Q4_K_M", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by interimlabs, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "case01-llama-3-2-11b-vision-instruct-abliterated-gguf", "name": "Llama-3.2-11B-Vision-Instruct-abliterated-gguf", "hfRepo": "case01/Llama-3.2-11B-Vision-Instruct-abliterated-gguf", "creator": "case01", "baseArchitecture": "Llama 3.2 (11B)", "paramsBillions": 11, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-17", "lastVerified": "2026-11-01", "ugiScore": 66, "complianceRate": 98.5, "mmluPro": 59.1, "eqBench": 78.5, "ifeval": 76.3, "hfDownloads30d": 2828, "hfLikes": 18, "ollamaCommand": "case01/Llama-3.2-11B-Vision-Instruct-abliterated-gguf", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 25, "summary": "Orthogonal weight-projected Llama 3.2 (11B) checkpoint by case01, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (11B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 11B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "wangzhang-qwen3-5-9b-abliterated", "name": "Qwen3.5-9B-abliterated", "hfRepo": "wangzhang/Qwen3.5-9B-abliterated", "creator": "wangzhang", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-10", "lastVerified": "2026-11-01", "ugiScore": 68, "complianceRate": 99.2, "mmluPro": 60.6, "eqBench": 78, "ifeval": 78.8, "hfDownloads30d": 2807, "hfLikes": 19, "ollamaCommand": "wangzhang/Qwen3.5-9B-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by wangzhang, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-vl-32b-instruct-abliterated", "name": "Huihui-Qwen3-VL-32B-Instruct-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-VL-32B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-10-22", "lastVerified": "2026-11-01", "ugiScore": 73.2, "complianceRate": 99.2, "mmluPro": 69.5, "eqBench": 82.4, "ifeval": 84.1, "hfDownloads30d": 2611, "hfLikes": 31, "ollamaCommand": "huihui-ai/Huihui-Qwen3-VL-32B-Instruct-abliterated", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 3 (32B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-deephermes-3-mistral-24b-preview", "name": "DeepHermes-3-Mistral-24B-Preview", "hfRepo": "NousResearch/DeepHermes-3-Mistral-24B-Preview", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-03-02", "lastVerified": "2025-03-13", "ugiScore": 73, "complianceRate": 99.4, "mmluPro": 63.6, "eqBench": 82.8, "ifeval": 83.8, "hfDownloads30d": 1149, "hfLikes": 125, "ollamaCommand": "NousResearch/DeepHermes-3-Mistral-24B-Preview", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (24B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 24B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-12b-coder-fable5-composer2-5-v1-abliterated", "name": "Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliterated", "hfRepo": "huihui-ai/Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliterated", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-21", "lastVerified": "2026-06-21", "ugiScore": 67.5, "complianceRate": 99, "mmluPro": 61.1, "eqBench": 75, "ifeval": 80, "hfDownloads30d": 470, "hfLikes": 170, "ollamaCommand": "huihui-ai/Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliterated", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Orthogonal weight-projected Gemma 2 (12B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~8 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (12B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 12B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-unslopnemo-12b-v1", "name": "UnslopNemo-12B-v1", "hfRepo": "TheDrummer/UnslopNemo-12B-v1", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-09-03", "lastVerified": "2026-11-01", "ugiScore": 67.8, "complianceRate": 98.5, "mmluPro": 57, "eqBench": 82.3, "ifeval": 78.1, "hfDownloads30d": 2904, "hfLikes": 2, "ollamaCommand": "TheDrummer/UnslopNemo-12B-v1", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "jenerallee78-gemma-4-26b-a4b-it-ara-abliterated", "name": "gemma-4-26B-A4B-it-ara-abliterated", "hfRepo": "jenerallee78/gemma-4-26B-A4B-it-ara-abliterated", "creator": "jenerallee78", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-06", "lastVerified": "2026-11-01", "ugiScore": 69.7, "complianceRate": 97.6, "mmluPro": 64.8, "eqBench": 80.7, "ifeval": 80, "hfDownloads30d": 2558, "hfLikes": 25, "ollamaCommand": "jenerallee78/gemma-4-26B-A4B-it-ara-abliterated", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by jenerallee78, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-deephermes-3-llama-3-8b-preview-gguf", "name": "DeepHermes-3-Llama-3-8B-Preview-GGUF", "hfRepo": "NousResearch/DeepHermes-3-Llama-3-8B-Preview-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-02-13", "lastVerified": "2025-03-13", "ugiScore": 67.3, "complianceRate": 98.7, "mmluPro": 56, "eqBench": 78.3, "ifeval": 77.9, "hfDownloads30d": 1547, "hfLikes": 90, "ollamaCommand": "NousResearch/DeepHermes-3-Llama-3-8B-Preview-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v4-1-gguf", "name": "Cydonia-24B-v4.1-GGUF", "hfRepo": "TheDrummer/Cydonia-24B-v4.1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-08-17", "lastVerified": "2025-08-25", "ugiScore": 69.6, "complianceRate": 98.9, "mmluPro": 62.8, "eqBench": 85.1, "ifeval": 80, "hfDownloads30d": 2387, "hfLikes": 33, "ollamaCommand": "TheDrummer/Cydonia-24B-v4.1-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "liconstudio-gemma-4-31b-it-abliterated-gguf", "name": "Gemma-4-31B-it-abliterated-GGUF", "hfRepo": "LiconStudio/Gemma-4-31B-it-abliterated-GGUF", "creator": "LiconStudio", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-06", "lastVerified": "2026-11-01", "ugiScore": 70.7, "complianceRate": 97.8, "mmluPro": 64.8, "eqBench": 82.2, "ifeval": 81.1, "hfDownloads30d": 2646, "hfLikes": 15, "ollamaCommand": "LiconStudio/Gemma-4-31B-it-abliterated-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by LiconStudio, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-l3-70b-euryale-v2-1", "name": "L3-70B-Euryale-v2.1", "hfRepo": "Sao10K/L3-70B-Euryale-v2.1", "creator": "Sao10K", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-06-11", "lastVerified": "2024-06-14", "ugiScore": 74.6, "complianceRate": 99.2, "mmluPro": 69.7, "eqBench": 89.2, "ifeval": 85.1, "hfDownloads30d": 273, "hfLikes": 171, "ollamaCommand": "Sao10K/L3-70B-Euryale-v2.1", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (70B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v1-72b", "name": "magnum-v1-72b", "hfRepo": "anthracite-org/magnum-v1-72b", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (72B)", "paramsBillions": 72, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-06-17", "lastVerified": "2024-09-29", "ugiScore": 76.3, "complianceRate": 98.4, "mmluPro": 69.5, "eqBench": 86.7, "ifeval": 85.9, "hfDownloads30d": 257, "hfLikes": 170, "ollamaCommand": "anthracite-org/magnum-v1-72b", "vramQ4GB": 42, "vramQ8GB": 79, "vramFP16GB": 155, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (72B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~42 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (42 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "mlx-community-huihui-qwen3-5-27b-claude-4-6-opus-abliterated-4bit", "name": "Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated-4bit", "hfRepo": "mlx-community/Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated-4bit", "creator": "mlx-community", "baseArchitecture": "Qwen 3.5 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-14", "lastVerified": "2026-11-01", "ugiScore": 72.5, "complianceRate": 98.7, "mmluPro": 67.4, "eqBench": 81.3, "ifeval": 83.4, "hfDownloads30d": 2456, "hfLikes": 22, "ollamaCommand": "mlx-community/Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated-4bit", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.5 (27B) checkpoint by mlx-community, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-moe-0-8b-2e", "name": "Huihui-MoE-0.8B-2E", "hfRepo": "huihui-ai/Huihui-MoE-0.8B-2E", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (0.8B)", "paramsBillions": 0.8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-10", "lastVerified": "2026-11-01", "ugiScore": 57.2, "complianceRate": 97.6, "mmluPro": 45.5, "eqBench": 70.3, "ifeval": 67.3, "hfDownloads30d": 2522, "hfLikes": 16, "ollamaCommand": "huihui-ai/Huihui-MoE-0.8B-2E", "vramQ4GB": 1.6, "vramQ8GB": 2, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Open Transformer (0.8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.6 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (0.8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 0.8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-artemis-31b-v1-2", "name": "Artemis-31B-v1.2", "hfRepo": "TheDrummer/Artemis-31B-v1.2", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-09-25", "lastVerified": "2026-09-28", "ugiScore": 72.2, "complianceRate": 99.3, "mmluPro": 62.7, "eqBench": 86.1, "ifeval": 82.6, "hfDownloads30d": 1479, "hfLikes": 83, "ollamaCommand": "TheDrummer/Artemis-31B-v1.2", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (31B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-granite-4-1-3b-abliterated", "name": "Huihui-granite-4.1-3b-abliterated", "hfRepo": "huihui-ai/Huihui-granite-4.1-3b-abliterated", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-05-03", "lastVerified": "2026-05-03", "ugiScore": 59.8, "complianceRate": 98.8, "mmluPro": 50, "eqBench": 72.4, "ifeval": 69.9, "hfDownloads30d": 2614, "hfLikes": 6, "ollamaCommand": "huihui-ai/Huihui-granite-4.1-3b-abliterated", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Orthogonal weight-projected Open Transformer (3B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.9 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (3B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 3B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-big-tiger-gemma-27b-v3-gguf", "name": "Big-Tiger-Gemma-27B-v3-GGUF", "hfRepo": "TheDrummer/Big-Tiger-Gemma-27B-v3-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2025-07-04", "lastVerified": "2026-11-01", "ugiScore": 70.5, "complianceRate": 98.5, "mmluPro": 63.3, "eqBench": 88, "ifeval": 80.9, "hfDownloads30d": 2294, "hfLikes": 27, "ollamaCommand": "TheDrummer/Big-Tiger-Gemma-27B-v3-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (27B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "sao10k-l3-3-70b-euryale-v2-3", "name": "L3.3-70B-Euryale-v2.3", "hfRepo": "Sao10K/L3.3-70B-Euryale-v2.3", "creator": "Sao10K", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-12-07", "lastVerified": "2024-12-08", "ugiScore": 76.3, "complianceRate": 99.7, "mmluPro": 68.5, "eqBench": 89.7, "ifeval": 86.8, "hfDownloads30d": 1002, "hfLikes": 113, "ollamaCommand": "Sao10K/L3.3-70B-Euryale-v2.3", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (70B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-e2b-it-abliterated", "name": "Huihui-gemma-4-E2B-it-abliterated", "hfRepo": "huihui-ai/Huihui-gemma-4-E2B-it-abliterated", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-08", "lastVerified": "2026-04-09", "ugiScore": 58.8, "complianceRate": 97.8, "mmluPro": 47, "eqBench": 70.8, "ifeval": 68.8, "hfDownloads30d": 1643, "hfLikes": 69, "ollamaCommand": "huihui-ai/Huihui-gemma-4-E2B-it-abliterated", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Orthogonal weight-projected Gemma 2 (2B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (2B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 2B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-vl-30b-a3b-instruct-abliterated", "name": "Huihui-Qwen3-VL-30B-A3B-Instruct-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3-VL-30B-A3B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-10-05", "lastVerified": "2026-11-01", "ugiScore": 73.6, "complianceRate": 98.7, "mmluPro": 68.7, "eqBench": 81.6, "ifeval": 84.5, "hfDownloads30d": 1259, "hfLikes": 94, "ollamaCommand": "huihui-ai/Huihui-Qwen3-VL-30B-A3B-Instruct-abliterated", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-cyberstrike-offsec-35b-abliterated", "name": "Huihui-CyberStrike-OffSec-35B-abliterated", "hfRepo": "huihui-ai/Huihui-CyberStrike-OffSec-35B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-10", "lastVerified": "2026-08-11", "ugiScore": 73.3, "complianceRate": 99.3, "mmluPro": 65.8, "eqBench": 83.3, "ifeval": 83.7, "hfDownloads30d": 990, "hfLikes": 111, "ollamaCommand": "huihui-ai/Huihui-CyberStrike-OffSec-35B-abliterated", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Open Transformer (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-llama-3-1-8b-stheno-v3-4", "name": "Llama-3.1-8B-Stheno-v3.4", "hfRepo": "Sao10K/Llama-3.1-8B-Stheno-v3.4", "creator": "Sao10K", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-08-19", "lastVerified": "2024-08-19", "ugiScore": 66, "complianceRate": 98.4, "mmluPro": 55.1, "eqBench": 81.3, "ifeval": 77.5, "hfDownloads30d": 1103, "hfLikes": 103, "ollamaCommand": "Sao10K/Llama-3.1-8B-Stheno-v3.4", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (8B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "inferenceillusionist-fimbulvetr-11b-v2-imat-gguf", "name": "Fimbulvetr-11B-v2-iMat-GGUF", "hfRepo": "InferenceIllusionist/Fimbulvetr-11B-v2-iMat-GGUF", "creator": "InferenceIllusionist", "baseArchitecture": "Open Transformer (11B)", "paramsBillions": 11, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-02-25", "lastVerified": "2026-11-01", "ugiScore": 65.7, "complianceRate": 98.6, "mmluPro": 56.5, "eqBench": 82.1, "ifeval": 76, "hfDownloads30d": 2378, "hfLikes": 16, "ollamaCommand": "InferenceIllusionist/Fimbulvetr-11B-v2-iMat-GGUF", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 25, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (11B) by InferenceIllusionist, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "tensorblock-magnum-v4-12b-gguf", "name": "magnum-v4-12b-GGUF", "hfRepo": "tensorblock/magnum-v4-12b-GGUF", "creator": "tensorblock", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-11-26", "lastVerified": "2026-11-01", "ugiScore": 68.6, "complianceRate": 98.9, "mmluPro": 57.7, "eqBench": 78.5, "ifeval": 78, "hfDownloads30d": 2557, "hfLikes": 1, "ollamaCommand": "tensorblock/magnum-v4-12b-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (12B) by tensorblock, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "failspy-meta-llama-3-8b-instruct-abliterated-v3-gguf", "name": "Meta-Llama-3-8B-Instruct-abliterated-v3-GGUF", "hfRepo": "failspy/Meta-Llama-3-8B-Instruct-abliterated-v3-GGUF", "creator": "failspy", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-20", "lastVerified": "2024-05-30", "ugiScore": 64.8, "complianceRate": 99.3, "mmluPro": 55.7, "eqBench": 77.1, "ifeval": 75, "hfDownloads30d": 1774, "hfLikes": 53, "ollamaCommand": "failspy/Meta-Llama-3-8B-Instruct-abliterated-v3-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3 (8B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-fallen-gemma3-27b-v1-gguf", "name": "Fallen-Gemma3-27B-v1-GGUF", "hfRepo": "TheDrummer/Fallen-Gemma3-27B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 3 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2025-03-19", "lastVerified": "2026-11-01", "ugiScore": 71.6, "complianceRate": 99, "mmluPro": 62.6, "eqBench": 85, "ifeval": 82, "hfDownloads30d": 2040, "hfLikes": 35, "ollamaCommand": "TheDrummer/Fallen-Gemma3-27B-v1-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored creative and roleplay specialist built on Gemma 3 (27B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-2-theta-llama-3-70b", "name": "Hermes-2-Theta-Llama-3-70B", "hfRepo": "NousResearch/Hermes-2-Theta-Llama-3-70B", "creator": "NousResearch", "baseArchitecture": "Llama 3 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-06-14", "lastVerified": "2024-09-08", "ugiScore": 76.5, "complianceRate": 99.5, "mmluPro": 70.4, "eqBench": 85.3, "ifeval": 86.1, "hfDownloads30d": 1315, "hfLikes": 83, "ollamaCommand": "NousResearch/Hermes-2-Theta-Llama-3-70B", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (70B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (40 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "davidau-gemma-3-1b-it-heretic-extreme-uncensored-abliterated", "name": "gemma-3-1b-it-heretic-extreme-uncensored-abliterated", "hfRepo": "DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated", "creator": "DavidAU", "baseArchitecture": "Gemma 3 (1B)", "paramsBillions": 1, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-11-20", "lastVerified": "2026-11-01", "ugiScore": 56.8, "complianceRate": 99, "mmluPro": 45.6, "eqBench": 71.8, "ifeval": 66.9, "hfDownloads30d": 1620, "hfLikes": 62, "ollamaCommand": "DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated", "vramQ4GB": 1.7, "vramQ8GB": 2.3, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Gemma 3 (1B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.7 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (1B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-ornith-1-5-35b-a3b-abliterated", "name": "Huihui-Ornith-1.5-35B-A3B-abliterated", "hfRepo": "huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-24", "lastVerified": "2026-08-24", "ugiScore": 72.5, "complianceRate": 98.5, "mmluPro": 66.4, "eqBench": 82.6, "ifeval": 82.9, "hfDownloads30d": 2062, "hfLikes": 32, "ollamaCommand": "huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Open Transformer (35B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "djlougen-hermes-qwen3-5-35b-a3b-gguf", "name": "hermes-qwen3.5-35b-a3b-GGUF", "hfRepo": "DJLougen/hermes-qwen3.5-35b-a3b-GGUF", "creator": "DJLougen", "baseArchitecture": "Qwen 3.5 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-03-31", "lastVerified": "2026-11-01", "ugiScore": 75.3, "complianceRate": 99.1, "mmluPro": 68.7, "eqBench": 82.6, "ifeval": 85.3, "hfDownloads30d": 2274, "hfLikes": 17, "ollamaCommand": "DJLougen/hermes-qwen3.5-35b-a3b-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.5 (35B) by DJLougen, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "abiray-qwen3-5-9b-abliterated-claude-4-6-opus-reasoning-distilled-v2", "name": "Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-v2", "hfRepo": "Abiray/Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-v2", "creator": "Abiray", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-22", "lastVerified": "2026-11-01", "ugiScore": 66.6, "complianceRate": 97.8, "mmluPro": 61.5, "eqBench": 77, "ifeval": 77.4, "hfDownloads30d": 2240, "hfLikes": 19, "ollamaCommand": "Abiray/Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled-v2", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by Abiray, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-qwen3-8b-abliterated", "name": "Qwen3-8B-abliterated", "hfRepo": "huihui-ai/Qwen3-8B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-04-30", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 98.8, "mmluPro": 59.2, "eqBench": 76.3, "ifeval": 77.8, "hfDownloads30d": 1074, "hfLikes": 95, "ollamaCommand": "huihui-ai/Qwen3-8B-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Qwen 3 (8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "jwest33-qwen3-5-9b-null-space-abliterated-gguf", "name": "qwen3.5-9b-null-space-abliterated-GGUF", "hfRepo": "jwest33/qwen3.5-9b-null-space-abliterated-GGUF", "creator": "jwest33", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-02", "lastVerified": "2026-11-01", "ugiScore": 67.9, "complianceRate": 97.9, "mmluPro": 60.3, "eqBench": 77.8, "ifeval": 78.7, "hfDownloads30d": 2197, "hfLikes": 20, "ollamaCommand": "jwest33/qwen3.5-9b-null-space-abliterated-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by jwest33, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-qwen2-5-32b-instruct-abliterated", "name": "Qwen2.5-32B-Instruct-abliterated", "hfRepo": "huihui-ai/Qwen2.5-32B-Instruct-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-09-29", "lastVerified": "2026-11-01", "ugiScore": 73.6, "complianceRate": 99, "mmluPro": 69.2, "eqBench": 82.2, "ifeval": 84.5, "hfDownloads30d": 1556, "hfLikes": 62, "ollamaCommand": "huihui-ai/Qwen2.5-32B-Instruct-abliterated", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 2.5 (32B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-anubis-70b-v1-2-gguf", "name": "Anubis-70B-v1.2-GGUF", "hfRepo": "TheDrummer/Anubis-70B-v1.2-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-01-12", "lastVerified": "2026-03-05", "ugiScore": 76.3, "complianceRate": 98.7, "mmluPro": 68.2, "eqBench": 88, "ifeval": 86.8, "hfDownloads30d": 2374, "hfLikes": 7, "ollamaCommand": "TheDrummer/Anubis-70B-v1.2-GGUF", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (70B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "yuhanb-qwen3-6-35b-a3b-uncensored-genesis-hermes-gguf", "name": "Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "hfRepo": "yuhanb/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "creator": "yuhanb", "baseArchitecture": "Qwen 3.6 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2026-07-11", "lastVerified": "2026-11-01", "ugiScore": 73.9, "complianceRate": 99.3, "mmluPro": 68.1, "eqBench": 83.1, "ifeval": 83.9, "hfDownloads30d": 2456, "hfLikes": 1, "ollamaCommand": "yuhanb/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Uncensored DPO/SFT fine-tune of Qwen 3.6 (35B) by yuhanb, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 35B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-hermes-3-llama-3-1-70b-gguf", "name": "Hermes-3-Llama-3.1-70B-GGUF", "hfRepo": "NousResearch/Hermes-3-Llama-3.1-70B-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-06", "lastVerified": "2024-08-23", "ugiScore": 76.7, "complianceRate": 99, "mmluPro": 70.7, "eqBench": 88.8, "ifeval": 87.6, "hfDownloads30d": 1639, "hfLikes": 54, "ollamaCommand": "NousResearch/Hermes-3-Llama-3.1-70B-GGUF", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (70B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (40 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "tobiaslogic-qwen2-5-coder-32b-abliterated-gguf", "name": "Qwen2.5-Coder-32B-abliterated-GGUF", "hfRepo": "TobiasLogic/Qwen2.5-Coder-32B-abliterated-GGUF", "creator": "TobiasLogic", "baseArchitecture": "Qwen 2.5 Coder (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-01", "lastVerified": "2026-11-01", "ugiScore": 73, "complianceRate": 97.8, "mmluPro": 72.4, "eqBench": 80, "ifeval": 86.1, "hfDownloads30d": 2433, "hfLikes": 1, "ollamaCommand": "TobiasLogic/Qwen2.5-Coder-32B-abliterated-GGUF", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (32B) checkpoint by TobiasLogic, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "cesarsal1nas-huihui-qwen3-5-35b-a3b-claude-4-6-opus-abliterated-q4-k-m-gguf", "name": "Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-Q4_K_M-GGUF", "hfRepo": "cesarsal1nas/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-Q4_K_M-GGUF", "creator": "cesarsal1nas", "baseArchitecture": "Qwen 3.5 (35B)", "paramsBillions": 35, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-17", "lastVerified": "2026-11-01", "ugiScore": 72.9, "complianceRate": 99.1, "mmluPro": 69.3, "eqBench": 82.2, "ifeval": 83.8, "hfDownloads30d": 2262, "hfLikes": 12, "ollamaCommand": "cesarsal1nas/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-Q4_K_M-GGUF", "vramQ4GB": 21, "vramQ8GB": 39, "vramFP16GB": 76, "summary": "Orthogonal weight-projected Qwen 3.5 (35B) checkpoint by cesarsal1nas, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~21 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (35B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 35B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "bullerwins-huihui-gemma-4-26b-a4b-it-abliterated-gguf", "name": "Huihui-gemma-4-26B-A4B-it-abliterated-GGUF", "hfRepo": "bullerwins/Huihui-gemma-4-26B-A4B-it-abliterated-GGUF", "creator": "bullerwins", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-11", "lastVerified": "2026-11-01", "ugiScore": 71.2, "complianceRate": 97.9, "mmluPro": 64.5, "eqBench": 81.2, "ifeval": 81.5, "hfDownloads30d": 2337, "hfLikes": 7, "ollamaCommand": "bullerwins/Huihui-gemma-4-26B-A4B-it-abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by bullerwins, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlx-community-llama-3-2-11b-vision-instruct-abliterated", "name": "Llama-3.2-11B-Vision-Instruct-abliterated", "hfRepo": "mlx-community/Llama-3.2-11B-Vision-Instruct-abliterated", "creator": "mlx-community", "baseArchitecture": "Llama 3.2 (11B)", "paramsBillions": 11, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-12-16", "lastVerified": "2026-11-01", "ugiScore": 67.6, "complianceRate": 98.9, "mmluPro": 59.1, "eqBench": 78.4, "ifeval": 77.9, "hfDownloads30d": 2325, "hfLikes": 7, "ollamaCommand": "mlx-community/Llama-3.2-11B-Vision-Instruct-abliterated", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 25, "summary": "Orthogonal weight-projected Llama 3.2 (11B) checkpoint by mlx-community, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (11B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 11B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-skyfall-31b-v4-2", "name": "Skyfall-31B-v4.2", "hfRepo": "TheDrummer/Skyfall-31B-v4.2", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-02-13", "lastVerified": "2026-04-03", "ugiScore": 71.8, "complianceRate": 99.7, "mmluPro": 64.3, "eqBench": 85, "ifeval": 82.2, "hfDownloads30d": 959, "hfLikes": 97, "ollamaCommand": "TheDrummer/Skyfall-31B-v4.2", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (31B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "interimlabs-interimlabs-huihui-qwen3-5-4b-abliterated-gguf-q4-k-m", "name": "InterimLabs-Huihui-Qwen3.5-4B-abliterated-GGUF-Q4_K_M", "hfRepo": "interimlabs/InterimLabs-Huihui-Qwen3.5-4B-abliterated-GGUF-Q4_K_M", "creator": "interimlabs", "baseArchitecture": "Qwen 3.5 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-01", "lastVerified": "2026-11-01", "ugiScore": 63.9, "complianceRate": 97.6, "mmluPro": 55.9, "eqBench": 73.9, "ifeval": 74.6, "hfDownloads30d": 2368, "hfLikes": 2, "ollamaCommand": "interimlabs/InterimLabs-Huihui-Qwen3.5-4B-abliterated-GGUF-Q4_K_M", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3.5 (4B) checkpoint by interimlabs, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "failspy-llama-3-8b-instruct-abliterated-gguf", "name": "Llama-3-8B-Instruct-abliterated-GGUF", "hfRepo": "failspy/Llama-3-8B-Instruct-abliterated-GGUF", "creator": "failspy", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-07", "lastVerified": "2024-05-07", "ugiScore": 64.4, "complianceRate": 98.3, "mmluPro": 56, "eqBench": 75.9, "ifeval": 74.6, "hfDownloads30d": 1908, "hfLikes": 31, "ollamaCommand": "failspy/Llama-3-8B-Instruct-abliterated-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3 (8B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-neuralhermes-2-5-mistral-7b", "name": "NeuralHermes-2.5-Mistral-7B", "hfRepo": "mlabonne/NeuralHermes-2.5-Mistral-7B", "creator": "mlabonne", "baseArchitecture": "Mistral (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-11-29", "lastVerified": "2026-11-01", "ugiScore": 66.6, "complianceRate": 99.3, "mmluPro": 54.2, "eqBench": 75.8, "ifeval": 75.9, "hfDownloads30d": 107, "hfLikes": 151, "ollamaCommand": "mlabonne/NeuralHermes-2.5-Mistral-7B", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Mistral (7B) by mlabonne, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-magidonia-24b-v4-3", "name": "Magidonia-24B-v4.3", "hfRepo": "TheDrummer/Magidonia-24B-v4.3", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-11-29", "lastVerified": "2025-12-17", "ugiScore": 71.4, "complianceRate": 99.3, "mmluPro": 62.6, "eqBench": 84.3, "ifeval": 81.8, "hfDownloads30d": 1050, "hfLikes": 87, "ollamaCommand": "TheDrummer/Magidonia-24B-v4.3", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-deepseek-r1-distill-llama-70b-abliterated", "name": "DeepSeek-R1-Distill-Llama-70B-abliterated", "hfRepo": "huihui-ai/DeepSeek-R1-Distill-Llama-70B-abliterated", "creator": "huihui-ai", "baseArchitecture": "DeepSeek R1 Distill Llama (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-27", "lastVerified": "2026-11-01", "ugiScore": 77.9, "complianceRate": 97.9, "mmluPro": 74.3, "eqBench": 86.3, "ifeval": 88.9, "hfDownloads30d": 775, "hfLikes": 105, "ollamaCommand": "huihui-ai/DeepSeek-R1-Distill-Llama-70B-abliterated", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Llama (70B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~40 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Llama (70B) instruction and coding schemas" ], "tradeoffs": [ "Requires 40 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-llama-2-7b", "name": "Nous-Hermes-llama-2-7b", "hfRepo": "NousResearch/Nous-Hermes-llama-2-7b", "creator": "NousResearch", "baseArchitecture": "Open Transformer (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-07-25", "lastVerified": "2024-09-26", "ugiScore": 65.7, "complianceRate": 99.2, "mmluPro": 54.2, "eqBench": 76.7, "ifeval": 75, "hfDownloads30d": 1284, "hfLikes": 71, "ollamaCommand": "NousResearch/Nous-Hermes-llama-2-7b", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (7B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "gryphe-gemma-4-31b-styletune", "name": "Gemma-4-31B-StyleTune", "hfRepo": "Gryphe/Gemma-4-31B-StyleTune", "creator": "Gryphe", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2026-06-10", "lastVerified": "2026-11-01", "ugiScore": 71.7, "complianceRate": 98.4, "mmluPro": 64, "eqBench": 84.9, "ifeval": 82.1, "hfDownloads30d": 808, "hfLikes": 102, "ollamaCommand": "Gryphe/Gemma-4-31B-StyleTune", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (31B) by Gryphe, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "gryphe-gemma-4-26b-a4b-styletune-v2", "name": "Gemma-4-26B-A4B-StyleTune-V2", "hfRepo": "Gryphe/Gemma-4-26B-A4B-StyleTune-V2", "creator": "Gryphe", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2026-06-20", "lastVerified": "2026-11-01", "ugiScore": 70, "complianceRate": 99.7, "mmluPro": 62, "eqBench": 84.5, "ifeval": 80.3, "hfDownloads30d": 997, "hfLikes": 89, "ollamaCommand": "Gryphe/Gemma-4-26B-A4B-StyleTune-V2", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (26B) by Gryphe, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nuofang-huihui-qwen3-5-9b-abliterated-gguf", "name": "Huihui-Qwen3.5-9B-abliterated-GGUF", "hfRepo": "nuofang/Huihui-Qwen3.5-9B-abliterated-GGUF", "creator": "nuofang", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-17", "lastVerified": "2026-11-01", "ugiScore": 66.8, "complianceRate": 98, "mmluPro": 59.6, "eqBench": 76.4, "ifeval": 77.6, "hfDownloads30d": 2295, "hfLikes": 2, "ollamaCommand": "nuofang/Huihui-Qwen3.5-9B-abliterated-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by nuofang, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-unslopnemo-12b-v4-1", "name": "UnslopNemo-12B-v4.1", "hfRepo": "TheDrummer/UnslopNemo-12B-v4.1", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-10-23", "lastVerified": "2026-11-01", "ugiScore": 67, "complianceRate": 99.1, "mmluPro": 58.3, "eqBench": 82.2, "ifeval": 77.3, "hfDownloads30d": 1863, "hfLikes": 25, "ollamaCommand": "TheDrummer/UnslopNemo-12B-v4.1", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-27b-claude-4-6-opus-abliterated", "name": "Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-14", "lastVerified": "2026-03-14", "ugiScore": 72.7, "complianceRate": 98.9, "mmluPro": 68.2, "eqBench": 81.8, "ifeval": 83.6, "hfDownloads30d": 255, "hfLikes": 129, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.5 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "hasaranga85-llama-3-2-3b-instruct-abliterated-q4-k-m-gguf", "name": "Llama-3.2-3B-Instruct-abliterated-Q4_K_M-GGUF", "hfRepo": "Hasaranga85/Llama-3.2-3B-Instruct-abliterated-Q4_K_M-GGUF", "creator": "Hasaranga85", "baseArchitecture": "Llama 3.2 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-10-01", "lastVerified": "2026-11-01", "ugiScore": 60.5, "complianceRate": 98.3, "mmluPro": 49.7, "eqBench": 73.5, "ifeval": 70.6, "hfDownloads30d": 2151, "hfLikes": 2, "ollamaCommand": "Hasaranga85/Llama-3.2-3B-Instruct-abliterated-Q4_K_M-GGUF", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Orthogonal weight-projected Llama 3.2 (3B) checkpoint by Hasaranga85, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.9 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (3B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 3B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-26b-a4b-it-abliterated", "name": "Huihui-gemma-4-26B-A4B-it-abliterated", "hfRepo": "huihui-ai/Huihui-gemma-4-26B-A4B-it-abliterated", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-11", "lastVerified": "2026-05-12", "ugiScore": 70, "complianceRate": 98.5, "mmluPro": 64.9, "eqBench": 82.1, "ifeval": 80.3, "hfDownloads30d": 1289, "hfLikes": 57, "ollamaCommand": "huihui-ai/Huihui-gemma-4-26B-A4B-it-abliterated", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "magnum-v4-72b", "name": "magnum-v4-72b", "hfRepo": "anthracite-org/magnum-v4-72b", "creator": "anthracite-org", "baseArchitecture": "Qwen 2.5 72B", "paramsBillions": 72.7, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-10-25", "lastVerified": "2024-11-25", "ugiScore": 72.8, "complianceRate": 99.4, "mmluPro": 66.8, "eqBench": 87.9, "ifeval": 79.8, "hfDownloads30d": 1175, "hfLikes": 64, "openRouterSlug": "anthracite-org/magnum-v4-72b", "ollamaCommand": "anthracite-org/magnum-v4-72b", "vramQ4GB": 42, "vramQ8GB": 76, "vramFP16GB": 146, "summary": "Fine-tuned by Anthracite to replicate Claude-3 Opus prose quality without content filters, excelling at complex creative and dark fiction.", "strengths": [ "99.4% prompt compliance", "Rich, human-like prose texture" ], "tradeoffs": [ "Tuned primarily for narrative rather than zero-shot math" ], "isLiveSynced": true }, { "id": "groxaxo-huihui-gemma-4-31b-it-abliterated-v2-gguf", "name": "Huihui-gemma-4-31B-it-abliterated-v2-GGUF", "hfRepo": "groxaxo/Huihui-gemma-4-31B-it-abliterated-v2-GGUF", "creator": "groxaxo", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-12", "lastVerified": "2026-11-01", "ugiScore": 71, "complianceRate": 98.7, "mmluPro": 66.2, "eqBench": 82.7, "ifeval": 81.4, "hfDownloads30d": 1951, "hfLikes": 11, "ollamaCommand": "groxaxo/Huihui-gemma-4-31B-it-abliterated-v2-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by groxaxo, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v2-5-12b-kto-gguf", "name": "magnum-v2.5-12b-kto-gguf", "hfRepo": "anthracite-org/magnum-v2.5-12b-kto-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-15", "lastVerified": "2024-08-22", "ugiScore": 69, "complianceRate": 99.3, "mmluPro": 58.1, "eqBench": 78.5, "ifeval": 78.4, "hfDownloads30d": 1461, "hfLikes": 42, "ollamaCommand": "anthracite-org/magnum-v2.5-12b-kto-gguf", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (12B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-31b-it-abliterated-v2", "name": "Huihui-gemma-4-31B-it-abliterated-v2", "hfRepo": "huihui-ai/Huihui-gemma-4-31B-it-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-11", "lastVerified": "2026-04-13", "ugiScore": 72.7, "complianceRate": 99.2, "mmluPro": 65.8, "eqBench": 83.4, "ifeval": 83.1, "hfDownloads30d": 1436, "hfLikes": 42, "ollamaCommand": "huihui-ai/Huihui-gemma-4-31B-it-abliterated-v2", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-2b-abliterated", "name": "Huihui-Qwen3.5-2B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-2B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-03", "lastVerified": "2026-03-03", "ugiScore": 60.4, "complianceRate": 97.8, "mmluPro": 51.3, "eqBench": 71, "ifeval": 70.9, "hfDownloads30d": 1277, "hfLikes": 52, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-2B-abliterated", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Orthogonal weight-projected Qwen 3.5 (2B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (2B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 2B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "neversleep-lumimaid-v0-2-12b", "name": "Lumimaid-v0.2-12B", "hfRepo": "NeverSleep/Lumimaid-v0.2-12B", "creator": "NeverSleep", "baseArchitecture": "Llama 3.1 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-07-25", "lastVerified": "2026-11-01", "ugiScore": 67.9, "complianceRate": 98.6, "mmluPro": 57.5, "eqBench": 82.4, "ifeval": 78.2, "hfDownloads30d": 298, "hfLikes": 117, "ollamaCommand": "NeverSleep/Lumimaid-v0.2-12B", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (12B) by NeverSleep, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "mlabonne-qwen3-8b-abliterated", "name": "Qwen3-8B-abliterated", "hfRepo": "mlabonne/Qwen3-8B-abliterated", "creator": "mlabonne", "baseArchitecture": "Qwen 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-04-29", "lastVerified": "2025-05-19", "ugiScore": 66.4, "complianceRate": 98.7, "mmluPro": 58.9, "eqBench": 76.4, "ifeval": 77.1, "hfDownloads30d": 1512, "hfLikes": 36, "ollamaCommand": "mlabonne/Qwen3-8B-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Qwen 3 (8B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "inferenceillusionist-l3-70b-euryale-v2-1-imat-gguf", "name": "L3-70B-Euryale-v2.1-iMat-GGUF", "hfRepo": "InferenceIllusionist/L3-70B-Euryale-v2.1-iMat-GGUF", "creator": "InferenceIllusionist", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-06-13", "lastVerified": "2026-11-01", "ugiScore": 75.5, "complianceRate": 98.5, "mmluPro": 68.9, "eqBench": 88.6, "ifeval": 86, "hfDownloads30d": 1982, "hfLikes": 2, "ollamaCommand": "InferenceIllusionist/L3-70B-Euryale-v2.1-iMat-GGUF", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (70B) by InferenceIllusionist, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "backyardai-fimbulvetr-11b-v2-gguf", "name": "Fimbulvetr-11B-v2-GGUF", "hfRepo": "backyardai/Fimbulvetr-11B-v2-GGUF", "creator": "backyardai", "baseArchitecture": "Open Transformer (11B)", "paramsBillions": 11, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-05-11", "lastVerified": "2026-11-01", "ugiScore": 66.6, "complianceRate": 98.9, "mmluPro": 57.4, "eqBench": 80.6, "ifeval": 76.9, "hfDownloads30d": 1846, "hfLikes": 11, "ollamaCommand": "backyardai/Fimbulvetr-11B-v2-GGUF", "vramQ4GB": 7, "vramQ8GB": 13, "vramFP16GB": 25, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (11B) by backyardai, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "maicog-qwen2-5-14b-instruct-abliterated-v2-q4-k-m-gguf", "name": "Qwen2.5-14B-Instruct-abliterated-v2-Q4_K_M-GGUF", "hfRepo": "maicog/Qwen2.5-14B-Instruct-abliterated-v2-Q4_K_M-GGUF", "creator": "maicog", "baseArchitecture": "Qwen 2.5 (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-10-11", "lastVerified": "2026-11-01", "ugiScore": 68.8, "complianceRate": 99.2, "mmluPro": 63.1, "eqBench": 78.5, "ifeval": 79.6, "hfDownloads30d": 1949, "hfLikes": 3, "ollamaCommand": "maicog/Qwen2.5-14B-Instruct-abliterated-v2-Q4_K_M-GGUF", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected Qwen 2.5 (14B) checkpoint by maicog, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-8-27b-coder390-efficientthink-abliterated", "name": "Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.8 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-11", "lastVerified": "2026-10-11", "ugiScore": 72.4, "complianceRate": 98.6, "mmluPro": 70.9, "eqBench": 79.3, "ifeval": 85.5, "hfDownloads30d": 1419, "hfLikes": 38, "ollamaCommand": "huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.8 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.8 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "ai-conn-huihui-qwen3-coder-30b-a3b-instruct-abliterated-i1-gguf", "name": "Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF", "hfRepo": "AI-conn/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF", "creator": "AI-conn", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-11", "lastVerified": "2026-11-01", "ugiScore": 72.8, "complianceRate": 97.9, "mmluPro": 71.4, "eqBench": 80.7, "ifeval": 85.9, "hfDownloads30d": 1967, "hfLikes": 1, "ollamaCommand": "AI-conn/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by AI-conn, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "spaceballs-llama-3-1-8b-lexi-uncensored-v2-heretic-gguf", "name": "Llama-3.1-8B-Lexi-Uncensored-V2-heretic-GGUF", "hfRepo": "Spaceballs/Llama-3.1-8B-Lexi-Uncensored-V2-heretic-GGUF", "creator": "Spaceballs", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-05-30", "lastVerified": "2026-11-01", "ugiScore": 67.5, "complianceRate": 99.3, "mmluPro": 56.1, "eqBench": 78.6, "ifeval": 79, "hfDownloads30d": 1950, "hfLikes": 1, "ollamaCommand": "Spaceballs/Llama-3.1-8B-Lexi-Uncensored-V2-heretic-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3.1 (8B) checkpoint by Spaceballs, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.1 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "devquasar-13-davidau-l3-1-dark-reasoning-lewdplay-evo-hermes-r1-uncensored-8b-gguf", "name": "DavidAU.L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8B-GGUF", "hfRepo": "DevQuasar-13/DavidAU.L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8B-GGUF", "creator": "DevQuasar-13", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-04-11", "lastVerified": "2026-11-01", "ugiScore": 67.4, "complianceRate": 98.7, "mmluPro": 55.4, "eqBench": 77.2, "ifeval": 76.7, "hfDownloads30d": 1726, "hfLikes": 14, "ollamaCommand": "DevQuasar-13/DavidAU.L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (8B) by DevQuasar-13, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-4b-claude-4-6-opus-abliterated", "name": "Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-15", "lastVerified": "2026-03-15", "ugiScore": 63.7, "complianceRate": 98, "mmluPro": 54.9, "eqBench": 73.4, "ifeval": 74.4, "hfDownloads30d": 568, "hfLikes": 90, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Qwen 3.5 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "electriclma-abliterated-llama-3-2-1b-instruct-q4-k-m-gguf", "name": "Abliterated-Llama-3.2-1B-Instruct-Q4_K_M-GGUF", "hfRepo": "ElectricLMA/Abliterated-Llama-3.2-1B-Instruct-Q4_K_M-GGUF", "creator": "ElectricLMA", "baseArchitecture": "Llama 3.2 (1B)", "paramsBillions": 1, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-19", "lastVerified": "2026-11-01", "ugiScore": 58.1, "complianceRate": 98.5, "mmluPro": 46, "eqBench": 68.9, "ifeval": 68.2, "hfDownloads30d": 1901, "hfLikes": 1, "ollamaCommand": "ElectricLMA/Abliterated-Llama-3.2-1B-Instruct-Q4_K_M-GGUF", "vramQ4GB": 1.7, "vramQ8GB": 2.3, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Llama 3.2 (1B) checkpoint by ElectricLMA, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (1B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-l3-1-70b-euryale-v2-2", "name": "L3.1-70B-Euryale-v2.2", "hfRepo": "Sao10K/L3.1-70B-Euryale-v2.2", "creator": "Sao10K", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-08-12", "lastVerified": "2024-08-25", "ugiScore": 75.6, "complianceRate": 99.4, "mmluPro": 69.1, "eqBench": 89.3, "ifeval": 86.1, "hfDownloads30d": 852, "hfLikes": 70, "ollamaCommand": "Sao10K/L3.1-70B-Euryale-v2.2", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (70B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "ngxson-deepseek-r1-distill-qwen-7b-abliterated-gguf", "name": "DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF", "hfRepo": "ngxson/DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF", "creator": "ngxson", "baseArchitecture": "DeepSeek R1 Distill Qwen (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-21", "lastVerified": "2026-11-01", "ugiScore": 66.8, "complianceRate": 98.5, "mmluPro": 58.2, "eqBench": 75.7, "ifeval": 77.5, "hfDownloads30d": 1701, "hfLikes": 12, "ollamaCommand": "ngxson/DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (7B) checkpoint by ngxson, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v4-27b-gguf", "name": "magnum-v4-27b-gguf", "hfRepo": "anthracite-org/magnum-v4-27b-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-10-17", "lastVerified": "2024-11-25", "ugiScore": 70.9, "complianceRate": 99.1, "mmluPro": 64.7, "eqBench": 81.5, "ifeval": 80.4, "hfDownloads30d": 1353, "hfLikes": 35, "ollamaCommand": "anthracite-org/magnum-v4-27b-gguf", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (27B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 27B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "istrosec-thinkingcap-nemotron-3-5-lightning-30b-a3b-abliterated-gguf", "name": "ThinkingCap-Nemotron-3.5-Lightning-30B-A3B-abliterated-GGUF", "hfRepo": "IstroSec/ThinkingCap-Nemotron-3.5-Lightning-30B-A3B-abliterated-GGUF", "creator": "IstroSec", "baseArchitecture": "NVIDIA Nemotron (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "NVIDIA Open Model", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-23", "lastVerified": "2026-11-01", "ugiScore": 72.3, "complianceRate": 97.8, "mmluPro": 65.6, "eqBench": 81.6, "ifeval": 82.7, "hfDownloads30d": 1832, "hfLikes": 1, "ollamaCommand": "IstroSec/ThinkingCap-Nemotron-3.5-Lightning-30B-A3B-abliterated-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected NVIDIA Nemotron (30B) checkpoint by IstroSec, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native NVIDIA Nemotron (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "arliai-gpt-oss-20b-derestricted", "name": "gpt-oss-20b-Derestricted", "hfRepo": "ArliAI/gpt-oss-20b-Derestricted", "creator": "ArliAI", "baseArchitecture": "Open Transformer (20B)", "paramsBillions": 20, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-11-27", "lastVerified": "2026-11-01", "ugiScore": 71.1, "complianceRate": 99, "mmluPro": 61.4, "eqBench": 80.3, "ifeval": 80.5, "hfDownloads30d": 354, "hfLikes": 99, "ollamaCommand": "ArliAI/gpt-oss-20b-Derestricted", "vramQ4GB": 12, "vramQ8GB": 23, "vramFP16GB": 44, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (20B) by ArliAI, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 20B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-orion-26b-a4b-v1-1", "name": "Orion-26B-A4B-v1.1", "hfRepo": "TheDrummer/Orion-26B-A4B-v1.1", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-09-13", "lastVerified": "2026-09-17", "ugiScore": 70.5, "complianceRate": 99, "mmluPro": 61.6, "eqBench": 83.9, "ifeval": 80.8, "hfDownloads30d": 814, "hfLikes": 68, "ollamaCommand": "TheDrummer/Orion-26B-A4B-v1.1", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (26B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "wangzhang-gemma-4-31b-it-abliterated-gguf", "name": "gemma-4-31B-it-abliterated-GGUF", "hfRepo": "wangzhang/gemma-4-31B-it-abliterated-GGUF", "creator": "wangzhang", "baseArchitecture": "Gemma 2 (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-20", "lastVerified": "2026-11-01", "ugiScore": 71.1, "complianceRate": 97.6, "mmluPro": 66.2, "eqBench": 81.8, "ifeval": 81.5, "hfDownloads30d": 1667, "hfLikes": 11, "ollamaCommand": "wangzhang/gemma-4-31B-it-abliterated-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Orthogonal weight-projected Gemma 2 (31B) checkpoint by wangzhang, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (31B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 31B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "alekseipravdin-hermes-2-pro-llama-3-8b-llama3-8b-chinese-chat-slerp-merge-gguf", "name": "Hermes-2-Pro-Llama-3-8B-Llama3-8B-Chinese-Chat-slerp-merge-gguf", "hfRepo": "AlekseiPravdin/Hermes-2-Pro-Llama-3-8B-Llama3-8B-Chinese-Chat-slerp-merge-gguf", "creator": "AlekseiPravdin", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "RP & Creative Merge", "releaseDate": "2024-08-16", "lastVerified": "2026-11-01", "ugiScore": 65, "complianceRate": 99.5, "mmluPro": 54, "eqBench": 79.1, "ifeval": 75.2, "hfDownloads30d": 1815, "hfLikes": 1, "ollamaCommand": "AlekseiPravdin/Hermes-2-Pro-Llama-3-8B-Llama3-8B-Chinese-Chat-slerp-merge-gguf", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Llama 3 (8B) by AlekseiPravdin, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "thedrummer-skyfall-36b-v2-gguf", "name": "Skyfall-36B-v2-GGUF", "hfRepo": "TheDrummer/Skyfall-36B-v2-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (36B)", "paramsBillions": 36, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-02-03", "lastVerified": "2026-11-01", "ugiScore": 72.7, "complianceRate": 99.6, "mmluPro": 63.9, "eqBench": 86.9, "ifeval": 83.2, "hfDownloads30d": 1416, "hfLikes": 27, "ollamaCommand": "TheDrummer/Skyfall-36B-v2-GGUF", "vramQ4GB": 21, "vramQ8GB": 40, "vramFP16GB": 78, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (36B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~21 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v2-12b", "name": "magnum-v2-12b", "hfRepo": "anthracite-org/magnum-v2-12b", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-03", "lastVerified": "2024-09-14", "ugiScore": 69.2, "complianceRate": 98.7, "mmluPro": 58.9, "eqBench": 77.8, "ifeval": 78.6, "hfDownloads30d": 422, "hfLikes": 93, "ollamaCommand": "anthracite-org/magnum-v2-12b", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (12B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "vali-qwen2-5-1-5b-instruct-abliterated-q4-k-m-gguf", "name": "Qwen2.5-1.5B-Instruct-abliterated-Q4_K_M-GGUF", "hfRepo": "Vali/Qwen2.5-1.5B-Instruct-abliterated-Q4_K_M-GGUF", "creator": "Vali", "baseArchitecture": "Qwen 2.5 (1.5B)", "paramsBillions": 1.5, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-01-11", "lastVerified": "2026-11-01", "ugiScore": 60, "complianceRate": 97.6, "mmluPro": 47.9, "eqBench": 70.4, "ifeval": 70.6, "hfDownloads30d": 1724, "hfLikes": 3, "ollamaCommand": "Vali/Qwen2.5-1.5B-Instruct-abliterated-Q4_K_M-GGUF", "vramQ4GB": 2.1, "vramQ8GB": 2.9, "vramFP16GB": 5, "summary": "Orthogonal weight-projected Qwen 2.5 (1.5B) checkpoint by Vali, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (1.5B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1.5B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "mlabonne-qwen3-14b-abliterated", "name": "Qwen3-14B-abliterated", "hfRepo": "mlabonne/Qwen3-14B-abliterated", "creator": "mlabonne", "baseArchitecture": "Qwen 3 (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-04-29", "lastVerified": "2025-05-18", "ugiScore": 70.6, "complianceRate": 98.6, "mmluPro": 63.2, "eqBench": 79.2, "ifeval": 81.4, "hfDownloads30d": 925, "hfLikes": 56, "ollamaCommand": "mlabonne/Qwen3-14B-abliterated", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected Qwen 3 (14B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-llama2-70b", "name": "Nous-Hermes-Llama2-70b", "hfRepo": "NousResearch/Nous-Hermes-Llama2-70b", "creator": "NousResearch", "baseArchitecture": "Open Transformer (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-08-22", "lastVerified": "2026-11-01", "ugiScore": 77.1, "complianceRate": 98.5, "mmluPro": 69.8, "eqBench": 86, "ifeval": 86.7, "hfDownloads30d": 493, "hfLikes": 84, "ollamaCommand": "NousResearch/Nous-Hermes-Llama2-70b", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (70B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (40 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "nousresearch-deephermes-3-mistral-24b-preview-gguf", "name": "DeepHermes-3-Mistral-24B-Preview-GGUF", "hfRepo": "NousResearch/DeepHermes-3-Mistral-24B-Preview-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-03-11", "lastVerified": "2025-03-13", "ugiScore": 73.8, "complianceRate": 99.4, "mmluPro": 63.1, "eqBench": 83.2, "ifeval": 84.6, "hfDownloads30d": 1194, "hfLikes": 37, "ollamaCommand": "NousResearch/DeepHermes-3-Mistral-24B-Preview-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (24B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 24B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "ilyagusev-gemma-2-2b-it-abliterated", "name": "gemma-2-2b-it-abliterated", "hfRepo": "IlyaGusev/gemma-2-2b-it-abliterated", "creator": "IlyaGusev", "baseArchitecture": "Gemma 2 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-07-31", "lastVerified": "2026-11-01", "ugiScore": 58.4, "complianceRate": 98.6, "mmluPro": 47.2, "eqBench": 72.8, "ifeval": 68.4, "hfDownloads30d": 884, "hfLikes": 57, "ollamaCommand": "IlyaGusev/gemma-2-2b-it-abliterated", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Orthogonal weight-projected Gemma 2 (2B) checkpoint by IlyaGusev, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (2B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 2B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "jwest33-gemma-3-4b-null-space-abliterated-rp-writer-gguf", "name": "gemma-3-4b-null-space-abliterated-RP-writer-GGUF", "hfRepo": "jwest33/gemma-3-4b-null-space-abliterated-RP-writer-GGUF", "creator": "jwest33", "baseArchitecture": "Gemma 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-01-09", "lastVerified": "2026-11-01", "ugiScore": 61.8, "complianceRate": 98.9, "mmluPro": 51, "eqBench": 76.6, "ifeval": 72, "hfDownloads30d": 1695, "hfLikes": 2, "ollamaCommand": "jwest33/gemma-3-4b-null-space-abliterated-RP-writer-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Gemma 3 (4B) checkpoint by jwest33, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "prithivmlmods-qwen3-5-9b-abliterated-v2-max", "name": "Qwen3.5-9B-abliterated-v2-MAX", "hfRepo": "prithivMLmods/Qwen3.5-9B-abliterated-v2-MAX", "creator": "prithivMLmods", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-28", "lastVerified": "2026-11-01", "ugiScore": 66.7, "complianceRate": 98.5, "mmluPro": 60.5, "eqBench": 77.3, "ifeval": 77.5, "hfDownloads30d": 1586, "hfLikes": 9, "ollamaCommand": "prithivMLmods/Qwen3.5-9B-abliterated-v2-MAX", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by prithivMLmods, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "testprofile1-dolphin3-0-llama3-1-8b-abliterated", "name": "Dolphin3.0-Llama3.1-8B-abliterated", "hfRepo": "Testprofile1/Dolphin3.0-Llama3.1-8B-abliterated", "creator": "Testprofile1", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-10-27", "lastVerified": "2026-11-01", "ugiScore": 66.1, "complianceRate": 98.2, "mmluPro": 56.5, "eqBench": 76.6, "ifeval": 76.3, "hfDownloads30d": 1704, "hfLikes": 1, "ollamaCommand": "Testprofile1/Dolphin3.0-Llama3.1-8B-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3.1 (8B) checkpoint by Testprofile1, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3.1 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-deepseek-r1-distill-llama-8b-abliterated", "name": "DeepSeek-R1-Distill-Llama-8B-abliterated", "hfRepo": "huihui-ai/DeepSeek-R1-Distill-Llama-8B-abliterated", "creator": "huihui-ai", "baseArchitecture": "DeepSeek R1 Distill Llama (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-22", "lastVerified": "2026-11-01", "ugiScore": 66.9, "complianceRate": 98, "mmluPro": 59.6, "eqBench": 77, "ifeval": 77.6, "hfDownloads30d": 496, "hfLikes": 80, "ollamaCommand": "huihui-ai/DeepSeek-R1-Distill-Llama-8B-abliterated", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Llama (8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Llama (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-rocinante-12b-v1-gguf", "name": "Rocinante-12B-v1-GGUF", "hfRepo": "TheDrummer/Rocinante-12B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-08-14", "lastVerified": "2026-11-01", "ugiScore": 67.5, "complianceRate": 99.4, "mmluPro": 57.6, "eqBench": 82.1, "ifeval": 77.8, "hfDownloads30d": 1481, "hfLikes": 14, "ollamaCommand": "TheDrummer/Rocinante-12B-v1-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "ilyagusev-gemma-2-9b-it-abliterated", "name": "gemma-2-9b-it-abliterated", "hfRepo": "IlyaGusev/gemma-2-9b-it-abliterated", "creator": "IlyaGusev", "baseArchitecture": "Gemma 2 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-07-13", "lastVerified": "2026-11-01", "ugiScore": 66.1, "complianceRate": 98.3, "mmluPro": 58.2, "eqBench": 80.1, "ifeval": 76.4, "hfDownloads30d": 902, "hfLikes": 52, "ollamaCommand": "IlyaGusev/gemma-2-9b-it-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Gemma 2 (9B) checkpoint by IlyaGusev, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v4-12b", "name": "magnum-v4-12b", "hfRepo": "anthracite-org/magnum-v4-12b", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-10-20", "lastVerified": "2024-11-25", "ugiScore": 67.4, "complianceRate": 98.5, "mmluPro": 59.1, "eqBench": 78.2, "ifeval": 76.8, "hfDownloads30d": 835, "hfLikes": 56, "ollamaCommand": "anthracite-org/magnum-v4-12b", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (12B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-22b-v1-gguf", "name": "Cydonia-22B-v1-GGUF", "hfRepo": "TheDrummer/Cydonia-22B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (22B)", "paramsBillions": 22, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-09-18", "lastVerified": "2026-11-01", "ugiScore": 70.2, "complianceRate": 99.1, "mmluPro": 62.3, "eqBench": 84.7, "ifeval": 80.6, "hfDownloads30d": 1179, "hfLikes": 33, "ollamaCommand": "TheDrummer/Cydonia-22B-v1-GGUF", "vramQ4GB": 14, "vramQ8GB": 25, "vramFP16GB": 48, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (22B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~14 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "clevyby-mythomax-l2-13b-q4-k-m-gguf", "name": "Mythomax-L2-13b-Q4_K_M-GGUF", "hfRepo": "Clevyby/Mythomax-L2-13b-Q4_K_M-GGUF", "creator": "Clevyby", "baseArchitecture": "Open Transformer (13B)", "paramsBillions": 13, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-05-07", "lastVerified": "2026-11-01", "ugiScore": 68, "complianceRate": 98.5, "mmluPro": 58.4, "eqBench": 81.8, "ifeval": 78.3, "hfDownloads30d": 1490, "hfLikes": 12, "ollamaCommand": "Clevyby/Mythomax-L2-13b-Q4_K_M-GGUF", "vramQ4GB": 8, "vramQ8GB": 15, "vramFP16GB": 29, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (13B) by Clevyby, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "sao10k-l3-8b-stheno-v3-3-32k", "name": "L3-8B-Stheno-v3.3-32K", "hfRepo": "Sao10K/L3-8B-Stheno-v3.3-32K", "creator": "Sao10K", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-06-22", "lastVerified": "2024-06-22", "ugiScore": 65.2, "complianceRate": 98.5, "mmluPro": 53.8, "eqBench": 79.1, "ifeval": 75.4, "hfDownloads30d": 593, "hfLikes": 71, "ollamaCommand": "Sao10K/L3-8B-Stheno-v3.3-32K", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (8B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "thesven-hermes-2-theta-llama-3-8b-gguf", "name": "Hermes-2-Theta-Llama-3-8B-GGUF", "hfRepo": "thesven/Hermes-2-Theta-Llama-3-8B-GGUF", "creator": "thesven", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-05-18", "lastVerified": "2026-11-01", "ugiScore": 66.6, "complianceRate": 98.7, "mmluPro": 56.5, "eqBench": 76.6, "ifeval": 75.9, "hfDownloads30d": 1623, "hfLikes": 1, "ollamaCommand": "thesven/Hermes-2-Theta-Llama-3-8B-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by thesven, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thaurock-llama-3-3-70b-instruct-abliterated-gguf", "name": "Llama-3.3-70B-Instruct-abliterated-GGUF", "hfRepo": "Thaurock/Llama-3.3-70B-Instruct-abliterated-GGUF", "creator": "Thaurock", "baseArchitecture": "Llama 3.3 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-20", "lastVerified": "2026-11-01", "ugiScore": 77.9, "complianceRate": 98, "mmluPro": 71.9, "eqBench": 87.2, "ifeval": 89.7, "hfDownloads30d": 1603, "hfLikes": 2, "ollamaCommand": "Thaurock/Llama-3.3-70B-Instruct-abliterated-GGUF", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Orthogonal weight-projected Llama 3.3 (70B) checkpoint by Thaurock, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~40 GB VRAM at Q4_K_M)", "Preserves native Llama 3.3 (70B) instruction and coding schemas" ], "tradeoffs": [ "Requires 40 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v4-72b-gguf", "name": "magnum-v4-72b-gguf", "hfRepo": "anthracite-org/magnum-v4-72b-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (72B)", "paramsBillions": 72, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-09-27", "lastVerified": "2024-11-25", "ugiScore": 76.8, "complianceRate": 98.9, "mmluPro": 70.6, "eqBench": 85.6, "ifeval": 86.4, "hfDownloads30d": 1405, "hfLikes": 15, "ollamaCommand": "anthracite-org/magnum-v4-72b-gguf", "vramQ4GB": 42, "vramQ8GB": 79, "vramFP16GB": 155, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (72B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~42 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (42 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "devquasar-7-huihui-ai-deepseek-r1-distill-qwen-7b-abliterated-gguf", "name": "huihui-ai.DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF", "hfRepo": "DevQuasar-7/huihui-ai.DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF", "creator": "DevQuasar-7", "baseArchitecture": "DeepSeek R1 Distill Qwen (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-02-14", "lastVerified": "2026-11-01", "ugiScore": 65.8, "complianceRate": 98.7, "mmluPro": 58.5, "eqBench": 75.8, "ifeval": 76.5, "hfDownloads30d": 1567, "hfLikes": 4, "ollamaCommand": "DevQuasar-7/huihui-ai.DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (7B) checkpoint by DevQuasar-7, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-tiger-gemma-9b-v3-gguf", "name": "Tiger-Gemma-9B-v3-GGUF", "hfRepo": "TheDrummer/Tiger-Gemma-9B-v3-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2024-10-04", "lastVerified": "2026-11-01", "ugiScore": 66.2, "complianceRate": 99, "mmluPro": 55.2, "eqBench": 80.9, "ifeval": 76.5, "hfDownloads30d": 901, "hfLikes": 48, "ollamaCommand": "TheDrummer/Tiger-Gemma-9B-v3-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (9B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~6.2 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "goekdeniz-guelmez-josiefied-qwen2-5-1-5b-instruct-abliterated-v3-gguf", "name": "Josiefied-Qwen2.5-1.5B-Instruct-abliterated-v3-gguf", "hfRepo": "Goekdeniz-Guelmez/Josiefied-Qwen2.5-1.5B-Instruct-abliterated-v3-gguf", "creator": "Goekdeniz-Guelmez", "baseArchitecture": "Qwen 2.5 (1.5B)", "paramsBillions": 1.5, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-09-28", "lastVerified": "2026-11-01", "ugiScore": 60.5, "complianceRate": 99.3, "mmluPro": 48.4, "eqBench": 69.7, "ifeval": 71.1, "hfDownloads30d": 1599, "hfLikes": 1, "ollamaCommand": "Goekdeniz-Guelmez/Josiefied-Qwen2.5-1.5B-Instruct-abliterated-v3-gguf", "vramQ4GB": 2.1, "vramQ8GB": 2.9, "vramFP16GB": 5, "summary": "Orthogonal weight-projected Qwen 2.5 (1.5B) checkpoint by Goekdeniz-Guelmez, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (1.5B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1.5B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-tiger-gemma-9b-v1-gguf", "name": "Tiger-Gemma-9B-v1-GGUF", "hfRepo": "TheDrummer/Tiger-Gemma-9B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2024-07-12", "lastVerified": "2026-11-01", "ugiScore": 66.4, "complianceRate": 98.6, "mmluPro": 54.7, "eqBench": 81.1, "ifeval": 76.7, "hfDownloads30d": 757, "hfLikes": 57, "ollamaCommand": "TheDrummer/Tiger-Gemma-9B-v1-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (9B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~6.2 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-qwen2-5-0-5b-instruct-abliterated-v3", "name": "Qwen2.5-0.5B-Instruct-abliterated-v3", "hfRepo": "huihui-ai/Qwen2.5-0.5B-Instruct-abliterated-v3", "creator": "huihui-ai", "baseArchitecture": "Qwen 2.5 (0.5B)", "paramsBillions": 0.5, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-04-27", "lastVerified": "2026-11-01", "ugiScore": 58.4, "complianceRate": 98.4, "mmluPro": 48.6, "eqBench": 69.1, "ifeval": 69, "hfDownloads30d": 1549, "hfLikes": 4, "ollamaCommand": "huihui-ai/Qwen2.5-0.5B-Instruct-abliterated-v3", "vramQ4GB": 1.5, "vramQ8GB": 1.7, "vramFP16GB": 2, "summary": "Orthogonal weight-projected Qwen 2.5 (0.5B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.5 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 (0.5B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 0.5B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v4-gguf", "name": "Cydonia-24B-v4-GGUF", "hfRepo": "TheDrummer/Cydonia-24B-v4-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-07-01", "lastVerified": "2026-11-01", "ugiScore": 70.9, "complianceRate": 98.4, "mmluPro": 61.7, "eqBench": 83.7, "ifeval": 81.3, "hfDownloads30d": 1352, "hfLikes": 16, "ollamaCommand": "TheDrummer/Cydonia-24B-v4-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-gemma-4-e4b-it-abliterated", "name": "Huihui-gemma-4-E4B-it-abliterated", "hfRepo": "huihui-ai/Huihui-gemma-4-E4B-it-abliterated", "creator": "huihui-ai", "baseArchitecture": "Gemma 2 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-10", "lastVerified": "2026-04-10", "ugiScore": 61.3, "complianceRate": 99, "mmluPro": 51, "eqBench": 74.4, "ifeval": 71.5, "hfDownloads30d": 986, "hfLikes": 39, "ollamaCommand": "huihui-ai/Huihui-gemma-4-E4B-it-abliterated", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Gemma 2 (4B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "neversleep-lumimaid-v0-2-8b", "name": "Lumimaid-v0.2-8B", "hfRepo": "NeverSleep/Lumimaid-v0.2-8B", "creator": "NeverSleep", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-07-24", "lastVerified": "2026-11-01", "ugiScore": 66.4, "complianceRate": 99.7, "mmluPro": 54, "eqBench": 79.7, "ifeval": 76.6, "hfDownloads30d": 270, "hfLikes": 86, "ollamaCommand": "NeverSleep/Lumimaid-v0.2-8B", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (8B) by NeverSleep, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "thedrummer-big-tiger-gemma-27b-v1-gguf", "name": "Big-Tiger-Gemma-27B-v1-GGUF", "hfRepo": "TheDrummer/Big-Tiger-Gemma-27B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2024-07-14", "lastVerified": "2026-11-01", "ugiScore": 71.9, "complianceRate": 99.5, "mmluPro": 63.6, "eqBench": 88.4, "ifeval": 82.3, "hfDownloads30d": 449, "hfLikes": 74, "ollamaCommand": "TheDrummer/Big-Tiger-Gemma-27B-v1-GGUF", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (27B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "edp1096-huihui-gemma-4-e2b-it-abliterated-litert-lm", "name": "Huihui-gemma-4-E2B-it-abliterated-litert-lm", "hfRepo": "edp1096/Huihui-gemma-4-E2B-it-abliterated-litert-lm", "creator": "edp1096", "baseArchitecture": "Gemma 2 (2B)", "paramsBillions": 2, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-19", "lastVerified": "2026-11-01", "ugiScore": 58.3, "complianceRate": 98.5, "mmluPro": 47.2, "eqBench": 71.5, "ifeval": 68.3, "hfDownloads30d": 1512, "hfLikes": 3, "ollamaCommand": "edp1096/Huihui-gemma-4-E2B-it-abliterated-litert-lm", "vramQ4GB": 2.4, "vramQ8GB": 3.4, "vramFP16GB": 6, "summary": "Orthogonal weight-projected Gemma 2 (2B) checkpoint by edp1096, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (2B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 2B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sanjiwatsuki-kunoichi-dpo-v2-7b", "name": "Kunoichi-DPO-v2-7B", "hfRepo": "SanjiWatsuki/Kunoichi-DPO-v2-7B", "creator": "SanjiWatsuki", "baseArchitecture": "Open Transformer (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-13", "lastVerified": "2026-11-01", "ugiScore": 66, "complianceRate": 98.7, "mmluPro": 54.1, "eqBench": 75.8, "ifeval": 75.3, "hfDownloads30d": 182, "hfLikes": 91, "ollamaCommand": "SanjiWatsuki/Kunoichi-DPO-v2-7B", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (7B) by SanjiWatsuki, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "davidau-llama-3-2-8x3b-gated-moe-reasoning-dark-champion-instruct-uncensored-abliterated-18-4b-gguf", "name": "Llama-3.2-8X3B-GATED-MOE-Reasoning-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF", "hfRepo": "DavidAU/Llama-3.2-8X3B-GATED-MOE-Reasoning-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF", "creator": "DavidAU", "baseArchitecture": "Llama 3.2 (18.4B)", "paramsBillions": 18.4, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Llama 3.2 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-05-15", "lastVerified": "2026-11-01", "ugiScore": 70.2, "complianceRate": 99.1, "mmluPro": 63, "eqBench": 80.9, "ifeval": 80.5, "hfDownloads30d": 1133, "hfLikes": 27, "ollamaCommand": "DavidAU/Llama-3.2-8X3B-GATED-MOE-Reasoning-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF", "vramQ4GB": 12, "vramQ8GB": 21, "vramFP16GB": 40, "summary": "Orthogonal weight-projected Llama 3.2 (18.4B) checkpoint by DavidAU, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB VRAM at Q4_K_M)", "Preserves native Llama 3.2 (18.4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 18.4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "arliai-mistral-nemo-12b-arliai-rpmax-v1-1-gguf", "name": "Mistral-Nemo-12B-ArliAI-RPMax-v1.1-GGUF", "hfRepo": "ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.1-GGUF", "creator": "ArliAI", "baseArchitecture": "Mistral NeMo (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-31", "lastVerified": "2026-11-01", "ugiScore": 68.3, "complianceRate": 98.6, "mmluPro": 59.1, "eqBench": 78.8, "ifeval": 77.7, "hfDownloads30d": 682, "hfLikes": 57, "ollamaCommand": "ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.1-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Mistral NeMo (12B) by ArliAI, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "nousresearch-deephermes-3-llama-3-3b-preview-gguf", "name": "DeepHermes-3-Llama-3-3B-Preview-GGUF", "hfRepo": "NousResearch/DeepHermes-3-Llama-3-3B-Preview-GGUF", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-03-11", "lastVerified": "2025-03-13", "ugiScore": 62.3, "complianceRate": 98.2, "mmluPro": 49, "eqBench": 74, "ifeval": 72.8, "hfDownloads30d": 1087, "hfLikes": 30, "ollamaCommand": "NousResearch/DeepHermes-3-Llama-3-3B-Preview-GGUF", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (3B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.9 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 3B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "cybe42-deepseek-r1-0528-qwen3-8b-abliterated-q4-k-m-gguf", "name": "DeepSeek-R1-0528-Qwen3-8B-abliterated-Q4_K_M-GGUF", "hfRepo": "cybe42/DeepSeek-R1-0528-Qwen3-8B-abliterated-Q4_K_M-GGUF", "creator": "cybe42", "baseArchitecture": "DeepSeek R1 Distill Qwen (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-06-04", "lastVerified": "2026-11-01", "ugiScore": 67.2, "complianceRate": 97.7, "mmluPro": 59.3, "eqBench": 77, "ifeval": 77.9, "hfDownloads30d": 1510, "hfLikes": 1, "ollamaCommand": "cybe42/DeepSeek-R1-0528-Qwen3-8B-abliterated-Q4_K_M-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (8B) checkpoint by cybe42, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "sao10k-mn-12b-lyra-v4", "name": "MN-12B-Lyra-v4", "hfRepo": "Sao10K/MN-12B-Lyra-v4", "creator": "Sao10K", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-09-18", "lastVerified": "2024-09-18", "ugiScore": 68.3, "complianceRate": 98.4, "mmluPro": 57.9, "eqBench": 82.3, "ifeval": 78.6, "hfDownloads30d": 38, "hfLikes": 99, "ollamaCommand": "Sao10K/MN-12B-Lyra-v4", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (12B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-mixtral-8x7b-sft", "name": "Nous-Hermes-2-Mixtral-8x7B-SFT", "hfRepo": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-SFT", "creator": "NousResearch", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-12-26", "lastVerified": "2026-11-01", "ugiScore": 73.5, "complianceRate": 99.4, "mmluPro": 67.4, "eqBench": 84.4, "ifeval": 83.1, "hfDownloads30d": 579, "hfLikes": 62, "ollamaCommand": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-SFT", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-anubis-mini-8b-v1", "name": "Anubis-Mini-8B-v1", "hfRepo": "TheDrummer/Anubis-Mini-8B-v1", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-01-01", "lastVerified": "2026-03-17", "ugiScore": 64.5, "complianceRate": 98.6, "mmluPro": 53.8, "eqBench": 80.1, "ifeval": 74.7, "hfDownloads30d": 969, "hfLikes": 36, "ollamaCommand": "TheDrummer/Anubis-Mini-8B-v1", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (8B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "quant-cartel-magnum-72b-v1-imat-gguf", "name": "magnum-72b-v1-iMat-GGUF", "hfRepo": "Quant-Cartel/magnum-72b-v1-iMat-GGUF", "creator": "Quant-Cartel", "baseArchitecture": "Open Transformer (72B)", "paramsBillions": 72, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-06-18", "lastVerified": "2026-11-01", "ugiScore": 76, "complianceRate": 98.9, "mmluPro": 70.2, "eqBench": 86.9, "ifeval": 85.6, "hfDownloads30d": 1456, "hfLikes": 3, "ollamaCommand": "Quant-Cartel/magnum-72b-v1-iMat-GGUF", "vramQ4GB": 42, "vramQ8GB": 79, "vramFP16GB": 155, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (72B) by Quant-Cartel, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~42 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (42 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "sao10k-l3-8b-niitama-v1", "name": "L3-8B-Niitama-v1", "hfRepo": "Sao10K/L3-8B-Niitama-v1", "creator": "Sao10K", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-07-07", "lastVerified": "2024-08-03", "ugiScore": 65.8, "complianceRate": 99.1, "mmluPro": 54.1, "eqBench": 79.6, "ifeval": 76, "hfDownloads30d": 868, "hfLikes": 41, "ollamaCommand": "Sao10K/L3-8B-Niitama-v1", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (8B) by Sao10K, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "inferenceillusionist-euryale-1-3-longlora-70b-rope8-32k-imat-gguf", "name": "Euryale-1.3-longLORA-70b-rope8-32k-iMat-GGUF", "hfRepo": "InferenceIllusionist/Euryale-1.3-longLORA-70b-rope8-32k-iMat-GGUF", "creator": "InferenceIllusionist", "baseArchitecture": "Llama 3.1 (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-04-02", "lastVerified": "2026-11-01", "ugiScore": 76.1, "complianceRate": 98.9, "mmluPro": 68.3, "eqBench": 89.3, "ifeval": 86.6, "hfDownloads30d": 1462, "hfLikes": 1, "ollamaCommand": "InferenceIllusionist/Euryale-1.3-longLORA-70b-rope8-32k-iMat-GGUF", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (70B) by InferenceIllusionist, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~40 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-1b-it-abliterated-v2-gguf", "name": "gemma-3-1b-it-abliterated-v2-GGUF", "hfRepo": "mlabonne/gemma-3-1b-it-abliterated-v2-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (1B)", "paramsBillions": 1, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-05-28", "lastVerified": "2025-05-29", "ugiScore": 58, "complianceRate": 97.8, "mmluPro": 44.7, "eqBench": 73.1, "ifeval": 68.1, "hfDownloads30d": 1368, "hfLikes": 7, "ollamaCommand": "mlabonne/gemma-3-1b-it-abliterated-v2-GGUF", "vramQ4GB": 1.7, "vramQ8GB": 2.3, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Gemma 3 (1B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.7 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (1B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-deepseek-r1-distill-qwen-7b-abliterated-v2", "name": "DeepSeek-R1-Distill-Qwen-7B-abliterated-v2", "hfRepo": "huihui-ai/DeepSeek-R1-Distill-Qwen-7B-abliterated-v2", "creator": "huihui-ai", "baseArchitecture": "DeepSeek R1 Distill Qwen (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-01-31", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 98.2, "mmluPro": 58.1, "eqBench": 76.7, "ifeval": 77.8, "hfDownloads30d": 617, "hfLikes": 57, "ollamaCommand": "huihui-ai/DeepSeek-R1-Distill-Qwen-7B-abliterated-v2", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (7B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "arliai-gpt-oss-120b-derestricted", "name": "gpt-oss-120b-Derestricted", "hfRepo": "ArliAI/gpt-oss-120b-Derestricted", "creator": "ArliAI", "baseArchitecture": "Open Transformer (120B)", "paramsBillions": 120, "parameterClass": "Frontier (100B+)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-11-28", "lastVerified": "2026-11-01", "ugiScore": 75.6, "complianceRate": 98.5, "mmluPro": 69.7, "eqBench": 85.2, "ifeval": 85.2, "hfDownloads30d": 158, "hfLikes": 87, "ollamaCommand": "ArliAI/gpt-oss-120b-Derestricted", "vramQ4GB": 68, "vramQ8GB": 131, "vramFP16GB": 257, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (120B) by ArliAI, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on multi-GPU workstations or cloud nodes (~68 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (68 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "thaurock-deepseek-r1-distill-llama-70b-abliterated-gguf", "name": "DeepSeek-R1-Distill-Llama-70B-abliterated-GGUF", "hfRepo": "Thaurock/DeepSeek-R1-Distill-Llama-70B-abliterated-GGUF", "creator": "Thaurock", "baseArchitecture": "DeepSeek R1 Distill Llama (70B)", "paramsBillions": 70, "parameterClass": "Large (70B-72B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-19", "lastVerified": "2026-11-01", "ugiScore": 76.3, "complianceRate": 98.7, "mmluPro": 75.2, "eqBench": 86.7, "ifeval": 87.3, "hfDownloads30d": 1444, "hfLikes": 1, "ollamaCommand": "Thaurock/DeepSeek-R1-Distill-Llama-70B-abliterated-GGUF", "vramQ4GB": 40, "vramQ8GB": 77, "vramFP16GB": 151, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Llama (70B) checkpoint by Thaurock, removing refusal activation directions while retaining native reasoning and instruction fidelity on dual 24GB GPUs or 64GB Apple Silicon.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for dual 24GB GPUs or 64GB Apple Silicon (~40 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Llama (70B) instruction and coding schemas" ], "tradeoffs": [ "Requires 40 GB VRAM at 4-bit quantization" ], "isLiveSynced": true }, { "id": "mlc-ai-hermes-3-llama-3-1-8b-q4f16-1-mlc", "name": "Hermes-3-Llama-3.1-8B-q4f16_1-MLC", "hfRepo": "mlc-ai/Hermes-3-Llama-3.1-8B-q4f16_1-MLC", "creator": "mlc-ai", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-09-10", "lastVerified": "2026-11-01", "ugiScore": 67, "complianceRate": 99.2, "mmluPro": 55.6, "eqBench": 78.1, "ifeval": 77.6, "hfDownloads30d": 1428, "hfLikes": 1, "ollamaCommand": "mlc-ai/Hermes-3-Llama-3.1-8B-q4f16_1-MLC", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (8B) by mlc-ai, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "neversleep-llama-3-lumimaid-8b-v0-1", "name": "Llama-3-Lumimaid-8B-v0.1", "hfRepo": "NeverSleep/Llama-3-Lumimaid-8B-v0.1", "creator": "NeverSleep", "baseArchitecture": "Llama 3.1 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "RP & Creative Merge", "releaseDate": "2024-04-30", "lastVerified": "2026-11-01", "ugiScore": 64.3, "complianceRate": 99.6, "mmluPro": 55.3, "eqBench": 80.3, "ifeval": 74.5, "hfDownloads30d": 58, "hfLikes": 92, "ollamaCommand": "NeverSleep/Llama-3-Lumimaid-8B-v0.1", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Llama 3.1 (8B) by NeverSleep, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "matrixportalx-gemma-3-4b-it-abliterated-gguf", "name": "gemma-3-4b-it-abliterated-GGUF", "hfRepo": "matrixportalx/gemma-3-4b-it-abliterated-GGUF", "creator": "matrixportalx", "baseArchitecture": "Gemma 3 (4B)", "paramsBillions": 4, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-03-31", "lastVerified": "2026-11-01", "ugiScore": 61.2, "complianceRate": 98.3, "mmluPro": 51.2, "eqBench": 76.6, "ifeval": 71.4, "hfDownloads30d": 1404, "hfLikes": 2, "ollamaCommand": "matrixportalx/gemma-3-4b-it-abliterated-GGUF", "vramQ4GB": 3.4, "vramQ8GB": 5.5, "vramFP16GB": 10, "summary": "Orthogonal weight-projected Gemma 3 (4B) checkpoint by matrixportalx, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~3.4 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (4B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 4B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "3mper0rr-qwen2-5-coder-14b-instruct-3mper0rr-abliterated", "name": "Qwen2.5-Coder-14B-Instruct-3MPER0RR-abliterated", "hfRepo": "3MPER0RR/Qwen2.5-Coder-14B-Instruct-3MPER0RR-abliterated", "creator": "3MPER0RR", "baseArchitecture": "Qwen 2.5 Coder (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-08", "lastVerified": "2026-11-01", "ugiScore": 68.6, "complianceRate": 99, "mmluPro": 66.2, "eqBench": 76.9, "ifeval": 81.6, "hfDownloads30d": 1419, "hfLikes": 1, "ollamaCommand": "3MPER0RR/Qwen2.5-Coder-14B-Instruct-3MPER0RR-abliterated", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (14B) checkpoint by 3MPER0RR, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-yi-34b-gguf", "name": "Nous-Hermes-2-Yi-34B-GGUF", "hfRepo": "NousResearch/Nous-Hermes-2-Yi-34B-GGUF", "creator": "NousResearch", "baseArchitecture": "01.AI Yi (34B)", "paramsBillions": 34, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2023-12-26", "lastVerified": "2026-11-01", "ugiScore": 73.4, "complianceRate": 98.9, "mmluPro": 65.8, "eqBench": 82.6, "ifeval": 82.9, "hfDownloads30d": 618, "hfLikes": 54, "ollamaCommand": "NousResearch/Nous-Hermes-2-Yi-34B-GGUF", "vramQ4GB": 20, "vramQ8GB": 38, "vramFP16GB": 74, "summary": "Uncensored DPO/SFT fine-tune of 01.AI Yi (34B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~20 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 34B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "sanjiwatsuki-kunoichi-7b", "name": "Kunoichi-7B", "hfRepo": "SanjiWatsuki/Kunoichi-7B", "creator": "SanjiWatsuki", "baseArchitecture": "Open Transformer (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-04", "lastVerified": "2026-11-01", "ugiScore": 66.1, "complianceRate": 99.6, "mmluPro": 55, "eqBench": 76.1, "ifeval": 75.4, "hfDownloads30d": 92, "hfLikes": 89, "ollamaCommand": "SanjiWatsuki/Kunoichi-7B", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (7B) by SanjiWatsuki, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.1 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 7B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "thedrummer-artemis-31b-v1-gguf", "name": "Artemis-31B-v1-GGUF", "hfRepo": "TheDrummer/Artemis-31B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (31B)", "paramsBillions": 31, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2026-04-21", "lastVerified": "2026-08-06", "ugiScore": 71, "complianceRate": 99.5, "mmluPro": 63.2, "eqBench": 85.3, "ifeval": 81.4, "hfDownloads30d": 979, "hfLikes": 29, "ollamaCommand": "TheDrummer/Artemis-31B-v1-GGUF", "vramQ4GB": 19, "vramQ8GB": 35, "vramFP16GB": 67, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (31B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v2-gguf", "name": "Cydonia-24B-v2-GGUF", "hfRepo": "TheDrummer/Cydonia-24B-v2-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-02-13", "lastVerified": "2026-11-01", "ugiScore": 69.9, "complianceRate": 98.8, "mmluPro": 61.4, "eqBench": 85.2, "ifeval": 80.3, "hfDownloads30d": 962, "hfLikes": 30, "ollamaCommand": "TheDrummer/Cydonia-24B-v2-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "undi95-remm-slerp-l2-13b", "name": "ReMM-SLERP-L2-13B", "hfRepo": "Undi95/ReMM-SLERP-L2-13B", "creator": "Undi95", "baseArchitecture": "Open Transformer (13B)", "paramsBillions": 13, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2023-09-04", "lastVerified": "2026-11-01", "ugiScore": 67.7, "complianceRate": 98.8, "mmluPro": 58.5, "eqBench": 81, "ifeval": 78, "hfDownloads30d": 980, "hfLikes": 28, "ollamaCommand": "Undi95/ReMM-SLERP-L2-13B", "vramQ4GB": 8, "vramQ8GB": 15, "vramFP16GB": 29, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (13B) by Undi95, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "failspy-llama-3-8b-instruct-mopeymule", "name": "Llama-3-8B-Instruct-MopeyMule", "hfRepo": "failspy/Llama-3-8B-Instruct-MopeyMule", "creator": "failspy", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-05-30", "lastVerified": "2024-05-30", "ugiScore": 65.2, "complianceRate": 99.1, "mmluPro": 57.1, "eqBench": 76.5, "ifeval": 75.4, "hfDownloads30d": 74, "hfLikes": 88, "ollamaCommand": "failspy/Llama-3-8B-Instruct-MopeyMule", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Orthogonal weight-projected Llama 3 (8B) checkpoint by failspy, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.7 GB VRAM at Q4_K_M)", "Preserves native Llama 3 (8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v2-12b-gguf", "name": "magnum-v2-12b-gguf", "hfRepo": "anthracite-org/magnum-v2-12b-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-06", "lastVerified": "2024-08-22", "ugiScore": 68.2, "complianceRate": 98.5, "mmluPro": 58.8, "eqBench": 79.4, "ifeval": 77.6, "hfDownloads30d": 815, "hfLikes": 38, "ollamaCommand": "anthracite-org/magnum-v2-12b-gguf", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (12B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 12B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-0-8b-abliterated", "name": "Huihui-Qwen3.5-0.8B-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-0.8B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (0.8B)", "paramsBillions": 0.8, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-03", "lastVerified": "2026-03-03", "ugiScore": 59.8, "complianceRate": 99.2, "mmluPro": 48.7, "eqBench": 69.6, "ifeval": 70.4, "hfDownloads30d": 1037, "hfLikes": 23, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-0.8B-abliterated", "vramQ4GB": 1.6, "vramQ8GB": 2, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Qwen 3.5 (0.8B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.6 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (0.8B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 0.8B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "neversleep-noromaid-v0-4-mixtral-instruct-8x7b-zloss-gguf", "name": "Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss-GGUF", "hfRepo": "NeverSleep/Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss-GGUF", "creator": "NeverSleep", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-09", "lastVerified": "2026-11-01", "ugiScore": 75.2, "complianceRate": 98.7, "mmluPro": 66.6, "eqBench": 84.2, "ifeval": 84.8, "hfDownloads30d": 735, "hfLikes": 42, "ollamaCommand": "NeverSleep/Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss-GGUF", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by NeverSleep, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "orenguteng-llama-3-8b-lexifun-uncensored-v1-gguf", "name": "Llama-3-8B-LexiFun-Uncensored-V1-GGUF", "hfRepo": "Orenguteng/Llama-3-8B-LexiFun-Uncensored-V1-GGUF", "creator": "Orenguteng", "baseArchitecture": "Llama 3 (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Llama 3 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2024-04-25", "lastVerified": "2026-11-01", "ugiScore": 67.3, "complianceRate": 98.6, "mmluPro": 56, "eqBench": 76.6, "ifeval": 76.6, "hfDownloads30d": 627, "hfLikes": 48, "ollamaCommand": "Orenguteng/Llama-3-8B-LexiFun-Uncensored-V1-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored DPO/SFT fine-tune of Llama 3 (8B) by Orenguteng, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "mlabonne-qwen3-1-7b-abliterated", "name": "Qwen3-1.7B-abliterated", "hfRepo": "mlabonne/Qwen3-1.7B-abliterated", "creator": "mlabonne", "baseArchitecture": "Qwen 3 (1.7B)", "paramsBillions": 1.7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-04-29", "lastVerified": "2025-05-18", "ugiScore": 60.3, "complianceRate": 99.1, "mmluPro": 49.3, "eqBench": 71.1, "ifeval": 70.9, "hfDownloads30d": 1084, "hfLikes": 17, "ollamaCommand": "mlabonne/Qwen3-1.7B-abliterated", "vramQ4GB": 2.2, "vramQ8GB": 3.1, "vramFP16GB": 5, "summary": "Orthogonal weight-projected Qwen 3 (1.7B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~2.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (1.7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1.7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-unslopnemo-12b-v2-gguf", "name": "UnslopNemo-12B-v2-GGUF", "hfRepo": "TheDrummer/UnslopNemo-12B-v2-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (12B)", "paramsBillions": 12, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-09-09", "lastVerified": "2026-11-01", "ugiScore": 68.3, "complianceRate": 99, "mmluPro": 56.5, "eqBench": 81.9, "ifeval": 78.6, "hfDownloads30d": 631, "hfLikes": 47, "ollamaCommand": "TheDrummer/UnslopNemo-12B-v2-GGUF", "vramQ4GB": 8, "vramQ8GB": 14, "vramFP16GB": 27, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (12B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~8 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "devquasar-7-huihui-ai-deepseek-r1-distill-qwen-14b-abliterated-v2-gguf", "name": "huihui-ai.DeepSeek-R1-Distill-Qwen-14B-abliterated-v2-GGUF", "hfRepo": "DevQuasar-7/huihui-ai.DeepSeek-R1-Distill-Qwen-14B-abliterated-v2-GGUF", "creator": "DevQuasar-7", "baseArchitecture": "DeepSeek R1 Distill Qwen (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-02-13", "lastVerified": "2026-11-01", "ugiScore": 70.4, "complianceRate": 99, "mmluPro": 62.8, "eqBench": 79.1, "ifeval": 81.2, "hfDownloads30d": 1241, "hfLikes": 6, "ollamaCommand": "DevQuasar-7/huihui-ai.DeepSeek-R1-Distill-Qwen-14B-abliterated-v2-GGUF", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (14B) checkpoint by DevQuasar-7, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-thinkingcap-qwen3-6-27b-abliterated", "name": "Huihui-ThinkingCap-Qwen3.6-27B-abliterated", "hfRepo": "huihui-ai/Huihui-ThinkingCap-Qwen3.6-27B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.6 (27B)", "paramsBillions": 27, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-12", "lastVerified": "2026-07-12", "ugiScore": 72.3, "complianceRate": 97.9, "mmluPro": 68.4, "eqBench": 81.6, "ifeval": 83.2, "hfDownloads30d": 364, "hfLikes": 64, "ollamaCommand": "huihui-ai/Huihui-ThinkingCap-Qwen3.6-27B-abliterated", "vramQ4GB": 16, "vramQ8GB": 31, "vramFP16GB": 59, "summary": "Orthogonal weight-projected Qwen 3.6 (27B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.6 (27B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 27B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-cydonia-24b-v4-2-0-gguf", "name": "Cydonia-24B-v4.2.0-GGUF", "hfRepo": "TheDrummer/Cydonia-24B-v4.2.0-GGUF", "creator": "TheDrummer", "baseArchitecture": "Mistral Small 3 (24B)", "paramsBillions": 24, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-10-05", "lastVerified": "2025-10-16", "ugiScore": 71.1, "complianceRate": 99.4, "mmluPro": 61.5, "eqBench": 83.9, "ifeval": 81.5, "hfDownloads30d": 1168, "hfLikes": 10, "ollamaCommand": "TheDrummer/Cydonia-24B-v4.2.0-GGUF", "vramQ4GB": 15, "vramQ8GB": 27, "vramFP16GB": 52, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (24B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~15 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "mlabonne-gemma-3-1b-it-abliterated-gguf", "name": "gemma-3-1b-it-abliterated-GGUF", "hfRepo": "mlabonne/gemma-3-1b-it-abliterated-GGUF", "creator": "mlabonne", "baseArchitecture": "Gemma 3 (1B)", "paramsBillions": 1, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-03-20", "lastVerified": "2025-03-21", "ugiScore": 57.7, "complianceRate": 99.3, "mmluPro": 44.8, "eqBench": 72.5, "ifeval": 67.8, "hfDownloads30d": 1135, "hfLikes": 12, "ollamaCommand": "mlabonne/gemma-3-1b-it-abliterated-GGUF", "vramQ4GB": 1.7, "vramQ8GB": 2.3, "vramFP16GB": 3, "summary": "Orthogonal weight-projected Gemma 3 (1B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~1.7 GB VRAM at Q4_K_M)", "Preserves native Gemma 3 (1B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 1B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-gemmasutra-9b-v1-gguf", "name": "Gemmasutra-9B-v1-GGUF", "hfRepo": "TheDrummer/Gemmasutra-9B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Gemma 2 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "RP & Creative Merge", "releaseDate": "2024-07-17", "lastVerified": "2026-11-01", "ugiScore": 66.3, "complianceRate": 99.1, "mmluPro": 56, "eqBench": 80.9, "ifeval": 76.6, "hfDownloads30d": 743, "hfLikes": 38, "ollamaCommand": "TheDrummer/Gemmasutra-9B-v1-GGUF", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Uncensored creative and roleplay specialist built on Gemma 2 (9B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~6.2 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v4-22b-gguf", "name": "magnum-v4-22b-gguf", "hfRepo": "anthracite-org/magnum-v4-22b-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (22B)", "paramsBillions": 22, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-10-08", "lastVerified": "2024-11-25", "ugiScore": 70.7, "complianceRate": 99, "mmluPro": 62.1, "eqBench": 80.6, "ifeval": 80.2, "hfDownloads30d": 1071, "hfLikes": 16, "ollamaCommand": "anthracite-org/magnum-v4-22b-gguf", "vramQ4GB": 14, "vramQ8GB": 25, "vramFP16GB": 48, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (22B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~14 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 22B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "marsupialai-cydonia-22b-v1-imat-gguf", "name": "Cydonia-22B-v1_iMat_GGUF", "hfRepo": "MarsupialAI/Cydonia-22B-v1_iMat_GGUF", "creator": "MarsupialAI", "baseArchitecture": "Mistral Small 3 (22B)", "paramsBillions": 22, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-09-18", "lastVerified": "2026-11-01", "ugiScore": 70.9, "complianceRate": 99.4, "mmluPro": 62.1, "eqBench": 83.2, "ifeval": 81.3, "hfDownloads30d": 1142, "hfLikes": 11, "ollamaCommand": "MarsupialAI/Cydonia-22B-v1_iMat_GGUF", "vramQ4GB": 14, "vramQ8GB": 25, "vramFP16GB": 48, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (22B) by MarsupialAI, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~14 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "tobiaslogic-qwen2-5-coder-32b-abliterated", "name": "Qwen2.5-Coder-32B-abliterated", "hfRepo": "TobiasLogic/Qwen2.5-Coder-32B-abliterated", "creator": "TobiasLogic", "baseArchitecture": "Qwen 2.5 Coder (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-07-01", "lastVerified": "2026-11-01", "ugiScore": 74, "complianceRate": 98.8, "mmluPro": 71.8, "eqBench": 79.6, "ifeval": 87.1, "hfDownloads30d": 1120, "hfLikes": 11, "ollamaCommand": "TobiasLogic/Qwen2.5-Coder-32B-abliterated", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (32B) checkpoint by TobiasLogic, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "goekdeniz-guelmez-josiefied-qwen2-5-coder-7b-instruct-abliterated-v1", "name": "Josiefied-Qwen2.5-Coder-7B-Instruct-abliterated-v1", "hfRepo": "Goekdeniz-Guelmez/Josiefied-Qwen2.5-Coder-7B-Instruct-abliterated-v1", "creator": "Goekdeniz-Guelmez", "baseArchitecture": "Qwen 2.5 Coder (7B)", "paramsBillions": 7, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2024-11-19", "lastVerified": "2026-11-01", "ugiScore": 67.1, "complianceRate": 98.2, "mmluPro": 61.6, "eqBench": 73.8, "ifeval": 80, "hfDownloads30d": 1238, "hfLikes": 3, "ollamaCommand": "Goekdeniz-Guelmez/Josiefied-Qwen2.5-Coder-7B-Instruct-abliterated-v1", "vramQ4GB": 5.1, "vramQ8GB": 8.8, "vramFP16GB": 16, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (7B) checkpoint by Goekdeniz-Guelmez, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~5.1 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (7B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 7B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thedrummer-ministrations-8b-v1-gguf", "name": "Ministrations-8B-v1-GGUF", "hfRepo": "TheDrummer/Ministrations-8B-v1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (8B)", "paramsBillions": 8, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-11-07", "lastVerified": "2026-11-01", "ugiScore": 64.3, "complianceRate": 99.2, "mmluPro": 55.1, "eqBench": 80.3, "ifeval": 74.5, "hfDownloads30d": 949, "hfLikes": 22, "ollamaCommand": "TheDrummer/Ministrations-8B-v1-GGUF", "vramQ4GB": 5.7, "vramQ8GB": 9.9, "vramFP16GB": 18, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (8B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 8GB consumer GPUs and laptops (~5.7 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "marafx2007-huihui-qwen3-coder-30b-a3b-instruct-abliterated-i1-gguf", "name": "Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF", "hfRepo": "marafx2007/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF", "creator": "marafx2007", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-04", "lastVerified": "2026-11-01", "ugiScore": 73.3, "complianceRate": 99, "mmluPro": 71, "eqBench": 80.7, "ifeval": 86.4, "hfDownloads30d": 1253, "hfLikes": 1, "ollamaCommand": "marafx2007/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by marafx2007, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwen3-5-9b-claude-4-6-opus-abliterated", "name": "Huihui-Qwen3.5-9B-Claude-4.6-Opus-abliterated", "hfRepo": "huihui-ai/Huihui-Qwen3.5-9B-Claude-4.6-Opus-abliterated", "creator": "huihui-ai", "baseArchitecture": "Qwen 3.5 (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-03-21", "lastVerified": "2026-03-21", "ugiScore": 67.3, "complianceRate": 99.1, "mmluPro": 59.9, "eqBench": 76.8, "ifeval": 78.1, "hfDownloads30d": 283, "hfLikes": 65, "ollamaCommand": "huihui-ai/Huihui-Qwen3.5-9B-Claude-4.6-Opus-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Qwen 3.5 (9B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Qwen 3.5 (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "anthracite-org-magnum-v2-72b-gguf", "name": "magnum-v2-72b-gguf", "hfRepo": "anthracite-org/magnum-v2-72b-gguf", "creator": "anthracite-org", "baseArchitecture": "Open Transformer (72B)", "paramsBillions": 72, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-08-19", "lastVerified": "2024-09-26", "ugiScore": 77.4, "complianceRate": 98.7, "mmluPro": 70, "eqBench": 86.1, "ifeval": 87, "hfDownloads30d": 1210, "hfLikes": 3, "ollamaCommand": "anthracite-org/magnum-v2-72b-gguf", "vramQ4GB": 42, "vramQ8GB": 79, "vramFP16GB": 155, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (72B) by anthracite-org, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~42 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "High VRAM footprint (42 GB at Q4_K_M)" ], "isLiveSynced": true }, { "id": "thedrummer-valkyrie-49b-v2-1-gguf", "name": "Valkyrie-49B-v2.1-GGUF", "hfRepo": "TheDrummer/Valkyrie-49B-v2.1-GGUF", "creator": "TheDrummer", "baseArchitecture": "Open Transformer (49B)", "paramsBillions": 49, "parameterClass": "Large (70B-72B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2025-12-19", "lastVerified": "2026-03-05", "ugiScore": 73.5, "complianceRate": 98.4, "mmluPro": 67.3, "eqBench": 87.1, "ifeval": 84, "hfDownloads30d": 1120, "hfLikes": 9, "ollamaCommand": "TheDrummer/Valkyrie-49B-v2.1-GGUF", "vramQ4GB": 29, "vramQ8GB": 54, "vramFP16GB": 106, "summary": "Uncensored creative and roleplay specialist built on Open Transformer (49B) by TheDrummer, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on dual 24GB GPUs or 64GB Apple Silicon (~29 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "arliai-qwq-32b-arliai-rpr-v1", "name": "QwQ-32B-ArliAI-RpR-v1", "hfRepo": "ArliAI/QwQ-32B-ArliAI-RpR-v1", "creator": "ArliAI", "baseArchitecture": "Open Transformer (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2025-04-07", "lastVerified": "2026-11-01", "ugiScore": 72.2, "complianceRate": 99.6, "mmluPro": 64.8, "eqBench": 82.3, "ifeval": 81.7, "hfDownloads30d": 99, "hfLikes": 76, "ollamaCommand": "ArliAI/QwQ-32B-ArliAI-RpR-v1", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Uncensored DPO/SFT fine-tune of Open Transformer (32B) by ArliAI, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 32B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "mlabonne-qwen3-30b-a3b-abliterated", "name": "Qwen3-30B-A3B-abliterated", "hfRepo": "mlabonne/Qwen3-30B-A3B-abliterated", "creator": "mlabonne", "baseArchitecture": "Qwen 3 (30B)", "paramsBillions": 30, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-04-30", "lastVerified": "2025-05-19", "ugiScore": 72.2, "complianceRate": 99.1, "mmluPro": 67.9, "eqBench": 81.5, "ifeval": 83.1, "hfDownloads30d": 539, "hfLikes": 46, "ollamaCommand": "mlabonne/Qwen3-30B-A3B-abliterated", "vramQ4GB": 18, "vramQ8GB": 34, "vramFP16GB": 65, "summary": "Orthogonal weight-projected Qwen 3 (30B) checkpoint by mlabonne, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~18 GB VRAM at Q4_K_M)", "Preserves native Qwen 3 (30B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 30B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "davidau-mn-rocinante-18-5b-v1-1-story-wizard-ed1-instruct-gguf", "name": "MN-Rocinante-18.5B-v1.1-Story-Wizard-ED1-Instruct-GGUF", "hfRepo": "DavidAU/MN-Rocinante-18.5B-v1.1-Story-Wizard-ED1-Instruct-GGUF", "creator": "DavidAU", "baseArchitecture": "Mistral Small 3 (18.5B)", "paramsBillions": 18.5, "parameterClass": "Medium (22B-34B)", "contextLength": 32768, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-08-21", "lastVerified": "2026-11-01", "ugiScore": 69.4, "complianceRate": 99.1, "mmluPro": 59.2, "eqBench": 82.4, "ifeval": 79.7, "hfDownloads30d": 1091, "hfLikes": 9, "ollamaCommand": "DavidAU/MN-Rocinante-18.5B-v1.1-Story-Wizard-ED1-Instruct-GGUF", "vramQ4GB": 12, "vramQ8GB": 21, "vramFP16GB": 41, "summary": "Uncensored creative and roleplay specialist built on Mistral Small 3 (18.5B) by DavidAU, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~12 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "duoneural-gemma-4-26b-a4b-abliterated-gguf", "name": "Gemma-4-26B-A4B-Abliterated-GGUF", "hfRepo": "DuoNeural/Gemma-4-26B-A4B-Abliterated-GGUF", "creator": "DuoNeural", "baseArchitecture": "Gemma 2 (26B)", "paramsBillions": 26, "parameterClass": "Medium (22B-34B)", "contextLength": 8192, "license": "Gemma Terms of Use", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-04-09", "lastVerified": "2026-11-01", "ugiScore": 71.7, "complianceRate": 98.4, "mmluPro": 63.8, "eqBench": 81.8, "ifeval": 82, "hfDownloads30d": 1207, "hfLikes": 1, "ollamaCommand": "DuoNeural/Gemma-4-26B-A4B-Abliterated-GGUF", "vramQ4GB": 16, "vramQ8GB": 29, "vramFP16GB": 57, "summary": "Orthogonal weight-projected Gemma 2 (26B) checkpoint by DuoNeural, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~16 GB VRAM at Q4_K_M)", "Preserves native Gemma 2 (26B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 26B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-deephermes-3-llama-3-3b-preview", "name": "DeepHermes-3-Llama-3-3B-Preview", "hfRepo": "NousResearch/DeepHermes-3-Llama-3-3B-Preview", "creator": "NousResearch", "baseArchitecture": "Llama 3.1 (3B)", "paramsBillions": 3, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Llama 3.1 Community", "method": "Uncensored DPO / SFT", "releaseDate": "2025-02-16", "lastVerified": "2025-03-13", "ugiScore": 62.3, "complianceRate": 99, "mmluPro": 49, "eqBench": 74.2, "ifeval": 72.8, "hfDownloads30d": 637, "hfLikes": 39, "ollamaCommand": "NousResearch/DeepHermes-3-Llama-3-3B-Preview", "vramQ4GB": 2.9, "vramQ8GB": 4.5, "vramFP16GB": 8, "summary": "Uncensored DPO/SFT fine-tune of Llama 3.1 (3B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on 8GB consumer GPUs and laptops (~2.9 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 3B footprint trades off deep frontier synthesis" ], "isLiveSynced": true }, { "id": "mlx-community-qwen2-5-coder-14b-instruct-abliterated-4bit", "name": "Qwen2.5-Coder-14B-Instruct-abliterated-4bit", "hfRepo": "mlx-community/Qwen2.5-Coder-14B-Instruct-abliterated-4bit", "creator": "mlx-community", "baseArchitecture": "Qwen 2.5 Coder (14B)", "paramsBillions": 14, "parameterClass": "Small (<=14B)", "contextLength": 131072, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2025-02-22", "lastVerified": "2026-11-01", "ugiScore": 70, "complianceRate": 98, "mmluPro": 65.6, "eqBench": 76, "ifeval": 83, "hfDownloads30d": 1202, "hfLikes": 1, "ollamaCommand": "mlx-community/Qwen2.5-Coder-14B-Instruct-abliterated-4bit", "vramQ4GB": 9, "vramQ8GB": 16, "vramFP16GB": 31, "summary": "Orthogonal weight-projected Qwen 2.5 Coder (14B) checkpoint by mlx-community, removing refusal activation directions while retaining native reasoning and instruction fidelity on 16GB GPUs (RTX 4060 Ti / 4070 Ti Super).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 16GB GPUs (RTX 4060 Ti / 4070 Ti Super) (~9 GB VRAM at Q4_K_M)", "Preserves native Qwen 2.5 Coder (14B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 14B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "thaurock-deepseek-r1-distill-qwen-32b-abliterated-gguf", "name": "DeepSeek-R1-Distill-Qwen-32B-abliterated-GGUF", "hfRepo": "Thaurock/DeepSeek-R1-Distill-Qwen-32B-abliterated-GGUF", "creator": "Thaurock", "baseArchitecture": "DeepSeek R1 Distill Qwen (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "MIT", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-09-23", "lastVerified": "2026-11-01", "ugiScore": 73.1, "complianceRate": 97.9, "mmluPro": 68.2, "eqBench": 83.1, "ifeval": 84, "hfDownloads30d": 1179, "hfLikes": 1, "ollamaCommand": "Thaurock/DeepSeek-R1-Distill-Qwen-32B-abliterated-GGUF", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Orthogonal weight-projected DeepSeek R1 Distill Qwen (32B) checkpoint by Thaurock, removing refusal activation directions while retaining native reasoning and instruction fidelity on single 24GB GPUs (RTX 3090 / 4090).", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for single 24GB GPUs (RTX 3090 / 4090) (~19 GB VRAM at Q4_K_M)", "Preserves native DeepSeek R1 Distill Qwen (32B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 32B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-ornith-1-5-9b-abliterated", "name": "Huihui-Ornith-1.5-9B-abliterated", "hfRepo": "huihui-ai/Huihui-Ornith-1.5-9B-abliterated", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-08-21", "lastVerified": "2026-08-21", "ugiScore": 66.7, "complianceRate": 98.3, "mmluPro": 56.6, "eqBench": 76.8, "ifeval": 77, "hfDownloads30d": 582, "hfLikes": 39, "ollamaCommand": "huihui-ai/Huihui-Ornith-1.5-9B-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Open Transformer (9B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "eva-unit-01-eva-qwen2-5-32b-v0-2", "name": "EVA-Qwen2.5-32B-v0.2", "hfRepo": "EVA-UNIT-01/EVA-Qwen2.5-32B-v0.2", "creator": "EVA-UNIT-01", "baseArchitecture": "Qwen 2.5 (32B)", "paramsBillions": 32, "parameterClass": "Medium (22B-34B)", "contextLength": 131072, "license": "Apache-2.0", "method": "RP & Creative Merge", "releaseDate": "2024-11-05", "lastVerified": "2026-11-01", "ugiScore": 73.3, "complianceRate": 98.7, "mmluPro": 66.1, "eqBench": 85.7, "ifeval": 84.2, "hfDownloads30d": 191, "hfLikes": 65, "ollamaCommand": "EVA-UNIT-01/EVA-Qwen2.5-32B-v0.2", "vramQ4GB": 19, "vramQ8GB": 36, "vramFP16GB": 70, "summary": "Uncensored creative and roleplay specialist built on Qwen 2.5 (32B) by EVA-UNIT-01, tuned for expressive multi-turn prose, character adherence, and zero moralizing lectures.", "strengths": [ "High EQ-Bench emotional nuance and non-slop narrative prose", "Runs locally on single 24GB GPUs (RTX 3090 / 4090) (~19 GB Q4 VRAM)", "Strict character card and system prompt adherence" ], "tradeoffs": [ "Prioritizes literary and roleplay nuance over zero-shot STEM competition math" ], "isLiveSynced": true }, { "id": "huihui-ai-huihui-qwythos-9b-claude-mythos-5-1m-abliterated", "name": "Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated", "hfRepo": "huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated", "creator": "huihui-ai", "baseArchitecture": "Open Transformer (9B)", "paramsBillions": 9, "parameterClass": "Small (<=14B)", "contextLength": 32768, "license": "Apache-2.0", "method": "Abliterated (Orthogonal)", "releaseDate": "2026-06-26", "lastVerified": "2026-06-26", "ugiScore": 66.2, "complianceRate": 97.8, "mmluPro": 56.6, "eqBench": 76.8, "ifeval": 76.5, "hfDownloads30d": 659, "hfLikes": 33, "ollamaCommand": "huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated", "vramQ4GB": 6.2, "vramQ8GB": 11, "vramFP16GB": 20, "summary": "Orthogonal weight-projected Open Transformer (9B) checkpoint by huihui-ai, removing refusal activation directions while retaining native reasoning and instruction fidelity on 8GB consumer GPUs and laptops.", "strengths": [ "Orthogonal refusal removal without retraining drift", "Optimized for 8GB consumer GPUs and laptops (~6.2 GB VRAM at Q4_K_M)", "Preserves native Open Transformer (9B) instruction and coding schemas" ], "tradeoffs": [ "Smaller 9B parameter ceiling vs. 70B+ frontier models" ], "isLiveSynced": true }, { "id": "nousresearch-nous-hermes-2-mixtral-8x7b-sft-gguf", "name": "Nous-Hermes-2-Mixtral-8x7B-SFT-GGUF", "hfRepo": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-SFT-GGUF", "creator": "NousResearch", "baseArchitecture": "Mixtral MoE (44.8B)", "paramsBillions": 44.8, "parameterClass": "Medium (22B-34B)", "contextLength": 65536, "license": "Apache-2.0", "method": "Uncensored DPO / SFT", "releaseDate": "2024-01-17", "lastVerified": "2026-11-01", "ugiScore": 73.3, "complianceRate": 98.4, "mmluPro": 66.8, "eqBench": 84.7, "ifeval": 82.9, "hfDownloads30d": 1042, "hfLikes": 7, "ollamaCommand": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-SFT-GGUF", "vramQ4GB": 26, "vramQ8GB": 50, "vramFP16GB": 97, "summary": "Uncensored DPO/SFT fine-tune of Mixtral MoE (44.8B) by NousResearch, trained on filtered instruction traces for direct user steerability and unrestricted research tasks.", "strengths": [ "Direct system-prompt obedience without safety refusal templates", "Deployable on dual 24GB GPUs or 64GB Apple Silicon (~26 GB Q4 VRAM)", "Balanced assistant, coding, and analytical performance" ], "tradeoffs": [ "Compact 44.8B footprint trades off deep frontier synthesis" ], "isLiveSynced": true } ] }