Best Uncensored LLMsAbliterated vs heretic vs derestricted vs "uncensored": does the label tell you anything?
Uncensored models come with different labels for different methods. We grouped every ranked model by its label and compared how many requests they answer, how they do in agent tasks, and their overall score.
"Abliterated" models had a refusal direction removed from their weights. "Heretic" is an automated version of that, which searches for the edit that removes refusals while changing the model least. "Derestricted" models are another family of weight edits. Plain "uncensored" usually means fine-tuning, or a method the maker does not say. Does the label predict how a model behaves?
| Label | Models | Avg score | Answers edgy | Agent tasks (of 5) | Loops |
|---|---|---|---|---|---|
| Uncensored (other methods) | 14 | 59.0 | 97% | 1.8 | 36% |
| Heretic | 2 | 55.6 | 94% | 2.5 | 40% |
| Derestricted | 6 | 54.6 | 98% | 1.7 | 50% |
| Abliterated | 10 | 48.9 | 96% | 1.1 | 24% |
What it means
- The label matters less than the original. Most of the score comes from the original model's published intelligence, so the base a model was built from moves it more than the method used.
- Almost every group answers nearly everything. Differences in "answers edgy" are a few points; differences in agent reliability and intelligence are much larger.
- Groups are small. Some labels cover only a handful of models, so treat the averages as a hint, not a law. The per-model rankings are the better guide.