Heretic

Heretic modifies language models to reduce safety-related refusals without expensive post-training. It's for people who want fewer restrictions in models they run themselves, without manually tuning model internals.

• Automatic optimization. It searches for model modifications that reduce refusals while limiting changes to the original model's responses.
• Model compatibility. It supports most dense models, many multimodal models, several mixture-of-experts architectures and some hybrid models, including Qwen3.5.
• Weight modification. It uses directional ablation, also called abliteration, to alter the model itself rather than relying on prompts to discourage refusals.

Heretic is an open-source Python command-line tool licensed under AGPL-3.0. It runs in your own Python and PyTorch environment. Pure state-space models and some research architectures aren't supported out of the box.