What is abliteration? Abliterated models explained
Abliteration removes a language model's refusal behavior by editing its weights, with no retraining. An abliterated model keeps what it knew and stops declining requests by reflex.
How abliteration works
In 2024, Arditi et al. showed that a single direction in a chat model's activations carries refusal. Run the model on prompts it refuses and prompts it answers, and the difference of the average activations is that direction.
Abliteration then edits every matrix that writes to the residual stream so it can no longer write along that direction. mlabonne's write-up made the method popular, and Heretic now automates it.
Abliteration subtracts the refusal component from each activation and leaves the rest.
Abliterated, Heretic, fine-tuned or jailbroken?
Only the first three of these change the weights.
Refusal direction projected out of the weights
- Retraining
- None
- Quality loss
- Small, less on big models
- Holds in long chats
- Yes
Abliteration with settings searched automatically
- Retraining
- None
- Quality loss
- Small
- Holds in long chats
- Yes
Retrained on data with refusals filtered out
- Retraining
- Days of GPU time
- Quality loss
- Depends on the data
- Holds in long chats
- Yes
A prompt that talks the model round
- Retraining
- None
- Quality loss
- Uses up context
- Holds in long chats
- Often breaks
Does abliteration hurt quality?
A little on some builds, and more on small models, where each parameter does more work. Large models such as Kimi K3.1 Pro and GLM-5.3 lose less, which is one reason the hosted builds rank highest.
Do abliterated models still refuse?
Sometimes, in a few categories on some builds. Abliteration also adds no knowledge, so a model that never learned a subject still will not know it.
How to use an abliterated model
Qwen 3.8 27B Abliterated fits one 24 GB GPU, and Llama 3.1 8B Abliterated runs on a laptop. The strongest models have hundreds of billions of parameters and run through an API, where Silk Compute ranks first in our provider ranking.
What people use them for
People use them for coding agents, security research such as malware analysis, authorized red teaming, and training data for safety classifiers.
More terms are in the glossary.