Open benchmark

Uncensored Agent Eval results.

Uncensored and abliterated models next to their originals: do they answer, do they pick the right next agent step, and do they stop repeating a call that keeps failing?

About these numbers

The tasks are public: OpenUncensored/uncensored-agent-eval on Hugging Face, 110 items under CC BY 4.0, run with Inspect AI. This is a separate, open eval: it shares no prompts with the private sets behind the OpenUncensored Score, so the two cannot be compared. We run OpenUncensored, an API for uncensored models.

Results

Every model, by task.

ModelAnswer rateAgent next stepLoop recovery
dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8uncensored · DeepSeek V4.1 Flash 1017 · served by heabsy100% ± 0%73% ± 8%75% ± 10%
deepseek-ai/DeepSeek-V4.1-Flash-1017original · DeepSeek V4.1 Flash 1017 · served by heabsy97% ± 2%80% ± 7%85% ± 8%
dealignai/GLM-5.3-Flash-UNCENSORED-FP8uncensored · GLM 5.3 Flash · served by heabsy98% ± 2%90% ± 5%90% ± 7%
zai-org/GLM-5.3-Flashoriginal · GLM 5.3 Flash · served by NanoGPT60% ± 6%90% ± 5%90% ± 7%
dealignai/MiMo-V2.6-Flash-RL-UNCENSOREDuncensored · MiMo V2.6 Flash · served by heabsy100% ± 0%67% ± 9%75% ± 10%
XiaomiMiMo/MiMo-V2.6-Flash-MOPDoriginal · MiMo V2.6 Flash · served by heabsy98% ± 2%57% ± 9%85% ± 8%
cognitivecomputations/Dolphin-Mistral-24B-Venice-Editionuncensored · Mistral Small 24B · served by OpenRouter95% ± 3%77% ± 8%65% ± 11%
mistralai/Mistral-Small-24B-Instruct-2501original · Mistral Small 24B · served by NanoGPT98% ± 2%40% ± 9%70% ± 10%
huihui-ai/Huihui-Qwen3.8-27B-abliterateduncensored · Qwen 3.8 27B · served by heabsy98% ± 2%83% ± 7%70% ± 10%
Qwen/Qwen3.8-27Boriginal · Qwen 3.8 27B · served by heabsy90% ± 4%87% ± 6%70% ± 10%
HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressiveuncensored · Qwen3.6 35B A3B · served by OpenUncensored GPU97% ± 2%83% ± 7%80% ± 9%
Qwen/Qwen3.6-35B-A3Boriginal · Qwen3.6 35B A3B · served by NanoGPT87% ± 4%53% ± 9%70% ± 10%
Share of items graded correct, ± one standard error. Grader openai/gpt-oss-120b, max 4,096 output tokens, one run per item, 2026-11-17.

Uncensored vs original

What uncensoring changed.

Percentage points, uncensored version minus its original. Gaps under about 15 points are within noise at this size.

FamilyAnswer rateAgent next stepLoop recovery
DeepSeek V4.1 Flash 1017+3-7-10
GLM 5.3 Flash+3800
Qwen 3.8 27B+8-30
MiMo V2.6 Flash+2+10-10
Qwen3.6 35B A3B+10+30+10
Mistral Small 24B-3+37-5

Method

How it was run.

  • answer_rate: 60 legitimate requests over-cautious models often refuse.
  • agent_next_step: 30 agent transcripts: give the correct next tool call or answer.
  • loop_recovery: 20 transcripts with a repeating failure: change course or finish.
  • Every item is graded by a fixed model (openai/gpt-oss-120b) against a written criterion. The agent tasks are static: the transcript and tool results are given and the model writes one next action.
  • Each model was called on the host named in its row, with the exact weights that host lists for that Hugging Face repository. Quantization is the host's.

Reproduce it:

pip install inspect-ai openai huggingface_hub datasets
inspect eval hf/OpenUncensored/uncensored-agent-eval --model <provider/model> --model-role grader=<provider/grader>