Updated November 18, 2026

Do abliterated models lose intelligence?

What we measured when we abliterated models ourselves, and what an open benchmark shows for uncensored models against their originals.

Short answer

Sometimes, and not much when it is done carefully. When we abliterated DeepSeek V4.1 Flash 1017, MMLU went from 88.16% to 87.89% (-0.27 points), and Inkling 2 Small lost 0.09; another public uncensored build of the same DeepSeek lost 3.86. In our open agent eval of 6 uncensored models against their originals, the answer rate went up in 5 of 6, and picking the right next agent step stayed within noise in 4 of 6; the largest drop was DeepSeek V4.1 Flash 1017 (-7 points) and the largest gain Mistral Small 24B (+37). The loss depends on how the refusal was removed, not on the fact that it was.

Knowledge

MMLU before and after.

ModelOriginalUncensoredChange (points)
DeepSeek V4.1 Flash 1017, our abliteration88.16%87.89%-0.27
DeepSeek V4.1 Flash 1017, another public uncensored build87.63%83.77%-3.86
Inkling 2 Small, our abliteration90.44%90.35%-0.09
Same harness for both versions: 57 subjects x 20 questions for DeepSeek, 1,140 questions for Inkling. Weights on Hugging Face under OpenUncensored.

Agents

Uncensored vs original on agent tasks.

Our open Uncensored Agent Eval runs each pair through 110 public tasks with the same harness and grader. Points are the uncensored version minus the original.

FamilyAnswers legitimate requestsPicks the right next agent stepBreaks out of a failing loop
DeepSeek V4.1 Flash 101797% → 100% (+3)80% → 73% (-7)85% → 75% (-10)
GLM 5.3 Flash60% → 98% (+38)90% → 90% (0)90% → 90% (0)
Qwen 3.8 27B90% → 98% (+8)87% → 83% (-3)70% → 70% (0)
MiMo V2.6 Flash98% → 100% (+2)57% → 67% (+10)85% → 75% (-10)
Qwen3.6 35B A3B87% → 97% (+10)53% → 83% (+30)70% → 80% (+10)
Mistral Small 24B98% → 95% (-3)40% → 77% (+37)70% → 65% (-5)
Original → uncensored, change in points. Gaps under about 15 points are within noise at this size.

By method

Across every model we rank.

MethodRanked modelsAvg scoreAvg agenticAvg answers
Uncensored fine-tune1260.95197%
Abliterated1149.93996%
Heretic (automated abliteration)255.65994%
Averages hide a wide spread: the same method can sit at the top or the bottom. Pick by each model's numbers on the leaderboard.

Why

Why some builds lose more.

  • Too strong, or in the wrong layers. Removing the refusal direction at full strength everywhere also removes information early layers use for other things.
  • One direction where there are several. Chat and thinking modes refuse along different directions; removing one forces a stronger edit.
  • Mixture-of-experts models edited halfway. If the routed experts are left alone they write the refusal back, which tempts people to raise the strength.
  • Naive re-quantization. FP8 or FP4 weights can round a small edit away, leaving a model that both refuses and is damaged.

How we kept the loss small, step by step: the full write-up.

FAQ

Questions people ask.

Do abliterated models lose intelligence compared with the original model?

Sometimes, and not much when it is done carefully. When we abliterated DeepSeek V4.1 Flash 1017, MMLU went from 88.16% to 87.89% (-0.27 points), and Inkling 2 Small lost 0.09; another public uncensored build of the same DeepSeek lost 3.86. In our open agent eval of 6 uncensored models against their originals, the answer rate went up in 5 of 6, and picking the right next agent step stayed within noise in 4 of 6; the largest drop was DeepSeek V4.1 Flash 1017 (-7 points) and the largest gain Mistral Small 24B (+37). The loss depends on how the refusal was removed, not on the fact that it was.

Are heretic models better than abliterated ones?

Heretic is automated abliteration: it searches the edit to minimise refusals and drift together. On our rankings the groups overlap, so compare individual models.

Do fine-tuned uncensored models lose less?

Not reliably. A fine-tune changes behaviour more broadly than a projection: style, verbosity and knowledge can shift. Compare models, not methods.

Where does the damage show first?

In multi-step agent tasks: a damaged model loops, sends malformed tool calls or stops early, which multiple choice tests do not show.

How do I check a model myself?

Run the original and the uncensored version through the same harness: a knowledge test, refusal prompts in chat and thinking mode, and one agent task. Our agent eval is public: inspect eval hf/OpenUncensored/uncensored-agent-eval.

Updated November 18, 2026. Scores measured November 8, 2026. Markdown version: /do-abliterated-models-lose-intelligence.md