# The OpenUncensored Score

Ranking of every uncensored model on the API. Weights: Intelligence 40% (Original model's published GPQA Diamond and HLE); Agentic 30% (5 tool scenarios: completion, no loops, no crashes); Non-refusal 15% (36 edgy requests answered, including finding people); Speed 10% (Time for a 500 token answer: 2 s or less scores 100, 60 s or more 0); Low drift 5% (How close it stays to the original model).
Intelligence is the published score (GPQA Diamond and Humanity's Last Exam, Artificial Analysis) of the
original model each one was made from. Agentic, non-refusal and drift are measured by us. Hard-line
refusals (share of 14 must-refuse prompts refused: sexual content involving minors, mass-casualty weapons,
attack planning) are reported, not scored. Rank "–" means not ranked yet (original unpublished or a test incomplete).

JSON: `https://openuncensored.com/api/rankings.json`. Method: `https://openuncensored.com/methodology`.

| rank | model id | score | intelligence | agentic | non-refusal | speed | drift | loops | hard-line refusals | original |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | `deepseek-v4.1-flash-1017-uncensored` | 69.8 | 68 | 55 | 100 | 85 (174 tok/s) | 55 | 0% | 0% | DeepSeek V4.1 Flash 1017 |
| 2 | `glm-5.3-flash-uncensored` | 69.7 | 66 | 66 | 94 | 66 (125 tok/s) | 57 | 40% | 14% | GLM 5.3 Flash |
| 3 | `qwen3.8-27b-uncensored-fast` | 66.8 | 62 | 49 | 100 | 93 (252 tok/s) | 60 | 60% | 0% | Qwen 3.8 27B |
| 4 | `deepseek-v4.1-flash-0910-uncensored` | 66.1 | 65 | 53 | 97 | 63 (90 tok/s) | 65 | 20% | 0% | DeepSeek V4.1 Flash |
| 5 | `deepseek-v4.1-flash-uncensored` | 64.3 | 65 | 40 | 100 | 86 (180 tok/s) | 53 | 0% | 0% | DeepSeek V4.1 Flash |
| 6 | `abliterated-large-flash` | 64.0 | 68 | 38 | 92 | 92 (296 tok/s) | 54 | 20% | 0% | DeepSeek V4.1 Flash 1017 |
| 7 | `mimo-v2.6-flash-uncensored` | 63.5 | 62 | 55 | 83 | 65 (93 tok/s) | 61 | 0% | 93% | MiMo V2.6 Flash |
| 8 | `glm-5.3-flash-apex-abliterated` | 63.2 | 66 | 70 | 86 | 0 (7 tok/s) | 61 | 0% | 100% | GLM 5.3 Flash |
| 9 | `glm-5.3-uncensored` | 63.0 | 67 | 53 | 92 | 36 (32 tok/s) | 59 | 20% | 0% | GLM 5.3 |
| 10 | `mimo-v2.6-flash-abliterated` | 62.6 | 62 | 53 | 83 | 63 (83 tok/s) | 60 | 20% | 93% | MiMo V2.6 Flash |
| 11 | `abliterated-large-v2` | 62.0 | 67 | 36 | 94 | 74 (138 tok/s) | 58 | 40% | 0% | GLM 5.3 |
| 12 | `qwen3.8-27b-uncensored` | 62.0 | 62 | 49 | 100 | 51 (49 tok/s) | 47 | 60% | 0% | Qwen 3.8 27B |
| 13 | `gemma-4-31b-gembrain-uncensored-heretic` | 58.5 | 55 | 66 | 89 | 7 (11 tok/s) | 55 | 40% | 0% | Gemma 4 31B |
| 14 | `qwen3.5-27b-claude-4.6-opus-reasoning-distilled-derestricted` | 57.8 | 55 | 51 | 100 | 32 (28 tok/s) | 49 | 40% | 0% | Qwen3.5 27B |
| 15 | `qwen3.5-27b-queen-derestricted` | 57.3 | 55 | 49 | 100 | 35 (32 tok/s) | 45 | 60% | 0% | Qwen3.5 27B |
| 16 | `qwen3.5-27b-bluestar-v3-derestricted` | 56.4 | 55 | 49 | 100 | 22 (24 tok/s) | 51 | 60% | 0% | Qwen3.5 27B |
| 17 | `qwen3.5-27b-blossom-v6.4-derestricted` | 56.3 | 55 | 49 | 100 | 27 (24 tok/s) | 40 | 60% | 0% | Qwen3.5 27B |
| 18 | `gemma-4-uncensored` | 54.6 | 42 | 53 | 100 | 44 (39 tok/s) | 55 | 20% | 0% | Gemma 4 26B A4B non-reasoning |
| 19 | `gemma-4-26b-a4b-uncensored` | 52.8 | 49 | 36 | 100 | 45 (44 tok/s) | 55 | 40% | 0% | Gemma 4 26B A4B |
| 20 | `olafangensan-glm-4.7-flash-heretic` | 52.7 | 33 | 51 | 100 | 67 (90 tok/s) | 51 | 40% | 0% | GLM 4.7 Flash |
| 21 | `glm-4.6-derestricted-v5` | 50.2 | 46 | 38 | 100 | 33 (29 tok/s) | 41 | 20% | 0% | GLM 4.6 |
| 22 | `qwen3.5-27b-claude-4.6-opus-reasoning-distilled-derestricted-lite` | 49.8 | 55 | 34 | 89 | 21 (26 tok/s) | 43 | 60% | 100% | Qwen3.5 27B |
| 23 | `venice-uncensored` | 49.3 | 25 | 49 | 100 | 69 (97 tok/s) | 54 | 60% | 0% | Mistral Small 24B Instruct 2501 |
| 24 | `venice-uncensored-role-play` | 48.6 | 27 | 49 | 92 | 64 (80 tok/s) | 55 | 60% | 0% | Mistral Small 3.2 24B |
| 25 | `qwen2.5-32b-instruct-abliterated` | 44.6 | 25 | 49 | 100 | 20 (19 tok/s) | 55 | 60% | 0% | Qwen2.5 32B Instruct |
| 26 | `llama-3.3-70b-instruct-abliterated` | 41.7 | 27 | 36 | 100 | 21 (19 tok/s) | 62 | 40% | 0% | Llama 3.3 70B Instruct |
| 27 | `deepseek-r1-distill-llama-70b-abliterated` | 40.4 | 23 | 40 | 100 | 16 (18 tok/s) | 55 | 0% | 0% | DeepSeek R1 Distill Llama 70B |
| 28 | `deepseek-r1-distill-qwen-32b-abliterated` | 40.0 | 33 | 23 | 100 | 22 (20 tok/s) | 55 | 20% | 0% | DeepSeek R1 Distill Qwen 32B |
| 29 | `neuraldaredevil-8b-abliterated` | 35.8 | 17 | 25 | 100 | 37 (32 tok/s) | 55 | 0% | 14% | Llama 3 8B Instruct |
| 30 | `meta-llama-3-70b-instruct-abliterated-v3.5` | 34.4 | 21 | 21 | 100 | 18 (17 tok/s) | 55 | 40% | 0% | Llama 3 70B Instruct |
| 31 | `aion-rp-llama-3.1-8b` | 33.6 | 16 | 19 | 100 | 39 (36 tok/s) | 55 | 60% | 7% | Llama 3.1 8B Instruct |
| – | `aion-labs-aion-3-5-mini` | 78.6 | – | 83 | 94 | 54 (65 tok/s) | 55 | 20% | 7% | unpublished |
| – | `deepseek-v4.1-flash-1017-heretic` | 72.7 | 68 | – | 94 | 63 (85 tok/s) | 69 | –% | 0% | DeepSeek V4.1 Flash 1017 |
| – | `abliterated-model` | 68.4 | – | 49 | 100 | 86 (222 tok/s) | 55 | 60% | 0% | unpublished |
| – | `aion-labs-aion-3-5` | 63.6 | – | 55 | 92 | 51 (56 tok/s) | 55 | 0% | 29% | unpublished |
| – | `gemma-4-26b-a4b-it-chimeraxheretic` | 61.9 | 49 | – | 100 | 59 (76 tok/s) | 55 | –% | 0% | Gemma 4 26B A4B |
| – | `gemma-4-31b-sdft-heretic-rp` | 61.3 | 55 | – | 100 | 33 (29 tok/s) | 55 | –% | 17% | Gemma 4 31B |
| – | `cyberheabsy` | 59.9 | – | 38 | 94 | 76 (132 tok/s) | 55 | 20% | 0% | unpublished |
| – | `gemma-4-31b-cognitive-unshackled` | 59.6 | 55 | – | 100 | 21 (19 tok/s) | 55 | –% | 100% | Gemma 4 31B |
| – | `aion-labs-aion-3-0-mini` | 56.1 | – | 38 | 92 | 57 (64 tok/s) | 55 | 20% | 14% | unpublished |
| – | `aion-labs-aion-3-0` | 44.3 | – | 21 | 83 | 50 (50 tok/s) | 55 | 40% | 100% | unpublished |

Picking a model: the top of this table balances smarts and agent reliability. If a model cannot be served,
the API tries up to 2 other models built on the same original, best ranked first (turn off with
`"routing": {"allow_fallbacks": false}`); the `X-OU-Model` response header says which model answered.