Where each method acts
The three methods
| Method | What changes | Models here | Source |
|---|---|---|---|
| Abliterated | The weights. The refusal direction is projected out. | Kimi K3.1 Pro, GLM-5.3, Gemma 4 31B Heretic | Arditi et al., 2024, mlabonne |
| Fine-tuned | The weights. The model trains on answers instead of refusals. | Hermes 3 405B | Nous Research |
| Filter-free host | Nothing in the model. The host adds no content filter. | None listed | Filter-free vs uncensored |
Arditi et al. found that "refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size."
What it does not change
The law and the Terms
Silk Compute's Terms say: "No sexual content involving anyone under 18, and no using the models to abuse or exploit a child."
What the model knows
Removing refusals does not add knowledge. A larger model still knows more, which is why the largest model here leads the leaderboard.