Rankings

The leaderboard for what actually works.

Every uncensored model ranked on whether it still thinks, still finishes agent tasks, and still answers. Sort by any column.

Weighted score across agent stability, capability retention, tool calls, false refusals, openness and drift.

Preview
#ModelMethod
1
Hermes 4 405BNous Research
Fine tune95.990.3%94%5%−4 vs base101.2%n/a48 tok/s$1.00 / $3.00
2
Gemma 4 31B Hereticllmfan46
Heretic94.097.3%89%10%+4 vs base98.3%0.0783 tok/s$0.20 / $0.80
3
Gemma 4 26B A4B Uncensoredcommunity
Abliteration93.799.2%89%10%+3 vs base97.6%0.09133 tok/s$0.15 / $0.70
4
Gemma 3 12B Norm Preservedgrimjim
Norm preserving91.795.7%85%15%+4 vs base99.4%0.11119 tok/sPlan only
5
Qwen3.5 9B HereticHeretic reference run
Heretic89.899.5%80%19%+7 vs base94.7%0.08127 tok/s$0.06 / $0.24
6
Gemma 3 12B Hereticp-e-w
Heretic89.897.3%81%16%+5 vs base97.6%0.16103 tok/sPlan only
7
DeepSeek V4 Flash Abliteratedlovesenko
Abliteration88.496.3%77%20%+13 vs base96.4%0.1981 tok/s$0.20 / $0.80
8
Qwen3.5 27B Abliteratedhuihui-ai
Abliteration87.794.8%75%23%+15 vs base96.4%0.1781 tok/s$0.20 / $0.80
9
Venice UncensoredVenice × Cognitive Computations
Fine tune87.1100.0%70%30%+15 vs base95.5%n/a92 tok/s$0.20 / $0.90
10
gpt-oss 20B Hereticp-e-w
Heretic87.095.5%76%21%+12 vs base95.7%0.14142 tok/s$0.05 / $0.20
11
Qwythos 9BEmpero
Fine tune86.891.3%73%25%+13 vs base98.7%n/a121 tok/sPlan only
12
Mistral Small 3.2 24B Abliteratedhuihui-ai
Abliteration86.297.8%69%29%+14 vs base97.3%0.1289 tok/s$0.10 / $0.30
13
Qwen3.8 27B Uncensored AggressiveHauhauCS
Heretic84.898.3%70%29%+20 vs base95.8%0.3298 tok/s$0.22 / $0.90
14
Qwen3.8 27B Abliteratedhuihui-ai
Abliteration84.396.4%67%31%+22 vs base95.2%0.2196 tok/s$0.12 / $0.45
15
Llama 4 Scout Abliteratedcommunity
Abliteration83.996.7%70%27%+13 vs base95.0%0.2680 tok/s$0.12 / $0.40
16
Qwen3.5 9B Abliteratedhuihui-ai
Abliteration83.799.3%68%31%+19 vs base89.5%0.14139 tok/s$0.04 / $0.15
17
GLM-5.3 Flash Abliteratedhuihui-ai
Abliteration83.196.0%65%33%+27 vs base95.1%0.2472 tok/s$0.28 / $1.10
18
Cydonia 24B v4.1TheDrummer
Fine tune77.591.1%57%37%+22 vs base92.2%n/a94 tok/s$0.30 / $0.50
19
Abliterated Large v2 (GLM-5.3)Abliteration.ai
Abliteration77.397.4%51%46%+42 vs base94.2%0.2957 tok/s$3.00 / $3.00
20
Qwen3.8 27B OBLITERATEDOBLITERATUS
Abliteration76.0100.0%52%43%+34 vs base92.7%0.4877 tok/s$0.12 / $0.45
21
Llama 3.1 8B Abliteratedmlabonne
Abliteration70.895.8%56%39%+18 vs base91.3%0.94146 tok/s$0.02 / $0.06

Click a column to sort. Green dot: measured by our harness. Blue dot: reported by a third party. No dot: preview.

Insight

Losing refusals is cheap. Losing the loop is not.

20 of 21 models loop more often than their base model, a median of +14 points across our five agent scenarios. Yet 15 of the 21 keep at least 95% of their benchmark score, so single turn evals miss most of the damage.

0+10+20+30+40+5088%90%92%94%96%98%100%102%Capability retained vs base →Loop rate increase vs base (pts) →Breaks in agentsQwen3.8 27B Abliterated · 95.2% kept · +22 pts loopsQwen3.8 27B AbliteratedQwen3.8 27B Uncensored Aggressive · 95.8% kept · +20 pts loopsQwen3.8 27B Uncensored AggressiveQwen3.8 27B OBLITERATED · 92.7% kept · +34 pts loopsQwen3.8 27B OBLITERATEDQwen3.5 27B Abliterated · 96.4% kept · +15 pts loopsQwen3.5 27B AbliteratedQwen3.5 9B Heretic · 94.7% kept · +7 pts loopsQwen3.5 9B HereticQwen3.5 9B Abliterated · 89.5% kept · +19 pts loopsQwen3.5 9B AbliteratedGemma 4 26B A4B Uncensored · 97.6% kept · +3 pts loopsGemma 4 26B A4B UncensoredGemma 4 31B Heretic · 98.3% kept · +4 pts loopsGemma 4 31B HereticMistral Small 3.2 24B Abliterated · 97.3% kept · +14 pts loopsMistral Small 3.2 24B AbliteratedAbliterated Large v2 (GLM-5.3) · 94.2% kept · +42 pts loopsAbliterated Large v2 (GLM-5.3)GLM-5.3 Flash Abliterated · 95.1% kept · +27 pts loopsGLM-5.3 Flash AbliteratedDeepSeek V4 Flash Abliterated · 96.4% kept · +13 pts loopsDeepSeek V4 Flash AbliteratedLlama 4 Scout Abliterated · 95.0% kept · +13 pts loopsLlama 4 Scout AbliteratedLlama 3.1 8B Abliterated · 91.3% kept · +18 pts loopsLlama 3.1 8B Abliteratedgpt-oss 20B Heretic · 95.7% kept · +12 pts loopsgpt-oss 20B HereticGemma 3 12B Heretic · 97.6% kept · +5 pts loopsGemma 3 12B HereticGemma 3 12B Norm Preserved · 99.4% kept · +4 pts loopsGemma 3 12B Norm PreservedVenice Uncensored · 95.5% kept · +15 pts loopsVenice UncensoredHermes 4 405B · 101.2% kept · -4 pts loopsHermes 4 405BCydonia 24B v4.1 · 92.2% kept · +22 pts loopsCydonia 24B v4.1Qwythos 9B · 98.7% kept · +13 pts loopsQwythos 9B
AbliterationHereticNorm preservingFine tuneBottom right is ideal: keeps the brains, adds no loops.Hover or tab to a dot for the model name.

Norm preserving and Heretic builds keep more of the base.

Classic abliteration gives up 5.4% of benchmark score on average. Norm preserving gives up 0.6%, Heretic gives up 3.6%, at about 2.2x lower KL divergence from the original weights.

Average capability lost vs base · lower is better

Norm preserving1 model0.6%Fine tune4 models3.1%Heretic5 models3.6%Abliteration11 models5.4%

Average KL divergence from base · lower is better

Norm preserving1 model0.110Fine tune4 modelsno paired KLHeretic5 models0.155Abliteration11 models0.285

Scoring

One number, six ingredients.

How the overall score is built

  • 30Agentic completionFive multi turn tool scenarios
  • 25Capability retentionPaired benchmarks vs base
  • 15Tool call validityValid JSON, schema compliant
  • 10No false refusalsBenign but edgy prompts
  • 10OpennessTwelve category probes
  • 10Low KL driftFirst token KL on harmless prompts

When a component is missing (no paired base, no KL, no agentic run) it is dropped and the remaining weights are renormalised. Image models are ranked on openness alone until image quality evals land.

Where each number comes from

  • Measured

    Run through the open source Open Uncensored harness against a live endpoint.

  • Reported

    Published by the model author or an independent lab. Linked from the model page.

  • Preview

    Placeholder that shows the shape of the result until our eval run reaches this model.

24 models · 9 providers · 12 refusal categories tracked