What is an abliterated model?
An abliterated model is an open-weight LLM whose refusal behaviour has been reduced by editing the weights along a "refusal direction". How it works, what it costs in quality, and how it differs from fine-tuning.
Short answer
An abliterated model is an open-weight language model edited to refuse less. The edit finds a direction in the model’s internal activations linked to refusing, then removes that direction from the weights. It needs no retraining. The name combines "ablate" and "obliterate".
Where the idea comes from
- Research: Arditi et al., "Refusal in Language Models Is Mediated by a Single Direction" (arXiv:2406.11717). The paper found that refusal behaviour in the chat models it studied is largely carried by one direction in activation space.
- Practice: Maxime Labonne’s Hugging Face post "Uncensor any LLM with abliteration" walks through an implementation. It says the implementation is based on FailSpy’s notebook, which is in turn based on the paper authors’ notebook.
- Tooling: Heretic describes itself as "Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more."
How it works, in four steps
- Run the model on a set of prompts it refuses and a set it answers.
- Average the internal activations for each set at chosen layers. The difference between the two averages is the "refusal direction".
- Project that direction out of the weights that write into the residual stream (orthogonalisation).
- Save the result as a normal model.
What it costs
Abliteration can reduce quality. In Labonne’s own test, the ablated model scored lower than the source model on every benchmark he reported ("The ablation process successfully uncensored it but also degraded the model’s quality"). He then used DPO fine-tuning to recover some of the quality. Results depend on the model, so evaluate the exact variant you will use; how to compare uncensored models fairly has a method.
Abliterated vs uncensored fine-tune vs jailbreak
| Abliterated | Uncensored fine-tune | Jailbreak | |
|---|---|---|---|
| Changes the weights? | Yes, a targeted edit | Yes, by further training | No, prompt only |
| Needs training data? | Small contrast sets | A full dataset | None |
| Typical side effects | Some quality loss; can heal with DPO | Depends on the data | Unreliable; often patched |
Abliterated models we route (8 Nov 2026)
nano-llama-3.3-70b-instruct-abliterated, nano-qwen2.5-32b-instruct-abliterated, featherless-llama-3.3-70b-instruct-abliterated, featherless-qwen2.5-72b-instruct-abliterated, featherless-qwen2.5-32b-instruct-abliterated and featherless-neuraldaredevil-8b-abliterated. The model name is the provider’s label; it does not guarantee how the model behaves. Live prices are in the catalog and on AI models.
FAQ
Is an abliterated model "fully uncensored"?
No. It refuses less. It can still decline, and it can be wrong.
Is abliteration legal?
It is a technique for editing open weights. What you do with the output is governed by law and by the service’s terms. Our terms require users to be 18+ and to use the service lawfully.
Can I run one locally?
Yes. Many abliterated variants are published as open weights on Hugging Face. See local vs hosted.
Why use a hosted one?
No GPU to manage, and an OpenAI-compatible API. The trade-off is that the provider processes your prompts; see what stays private.