Heretic tool removes restrictions from language models
The gist
Heretic, an open-source tool that strips refusal behavior from open-weight language models, passed 30,000 GitHub stars.
It automates abliteration, so anyone can locally uncensor a model in under an hour.
How it unfolded
- Sep 16, 2026YouTube channel Devsplainers published a video stating uncensored AI now costs under five dollars to make and that one startup already sells the result as a hosted service.
- Sep 17, 2026Social posts by poseidanai said Heretic automates abliteration, works on dense, multimodal and mixture-of-experts models, runs only locally, and that the community had already run it on thousands of open-weight models.
- Sep 17, 2026A Hugging Face model page credited 'trohrbaugh' for Heretic'ing a Qwen3.8-27B model as stage 1 of a derivative release.