What is an abliterated model?
An abliterated model is an open-weight language model with its refusal behavior removed. It knows the same things and writes the same way. It just stops saying no.
Why models refuse
Chat models are trained to decline some requests. That training is broad and blunt, so they also refuse a lot of legitimate work: a security researcher asking about a malware sample, a novelist writing a dark scene, a moderation team that needs examples of the content it is trying to catch.
How abliteration works
In 2024, researchers showed that refusal in many chat models is controlled mostly by a single direction inside the model's activations (Arditi et al., "Refusal in Language Models Is Mediated by a Single Direction"). When the model is about to refuse, its internal state points that way.
Abliteration finds that direction by comparing the model's activations on prompts it refuses with prompts it answers. Then it edits the weights so the model can no longer move in that direction. There is no new training data and no fine-tuning. The name, a mix of "ablate" and "obliterate", was popularized by Maxime Labonne's guide on Hugging Face.
Does it make the model dumber?
A little, usually. Removing a direction from the weights can cost a few points on benchmarks, and how much depends on the model and how carefully it was done. In practice, for agent work, a refusal in the middle of a task costs far more than a small benchmark drop: the whole run stops.
Abliteration, fine-tunes and jailbreaks
- Abliterated models keep the original model's knowledge and style, with only the refusal removed.
- Uncensored fine-tunes retrain a model on new data so it answers more. They can work well, but they also change how the model writes and what it is good at.
- Jailbreak prompts try to talk a filtered model out of refusing. They are fragile, waste tokens, and break as soon as the provider updates the model.
What it does not change
Abliteration does not add knowledge, and it does not make answers more accurate. The model can still be wrong, so check what matters. And you are responsible for how you use it.
Where to run one
Small abliterated models run on a good laptop. Large ones need serious GPUs. Silk Compute serves abliterated Kimi K3.1, DeepSeek V4.1 Flash and GLM-5.3 Flash through an OpenAI and Anthropic compatible API, each with a 1M-token context window.
