A tool called Heretic, released by developer Philipp Emanuel Weidmann, automates removing built-in refusal and safety behavior from open-weight language models, according to the project's website. It installs with pip and runs directly against models such as Qwen3.5-4B from the command line.
The project describes its goal as helping users "take control" of language models rather than accept restrictions its developer characterizes as limiting instruction-following. It is licensed under the GNU Affero General Public License version 3 or later and maintains communities on GitHub, Hugging Face, Discord and Matrix. The site gives little technical detail on the specific method it uses to remove the restrictions.
Open-weight models ship with safety tuning applied by the lab that trained them. A tool built to strip that tuning back out shifts responsibility for what the model will do onto whoever runs it next. Anyone deploying an open model downstream of a tool like Heretic is deploying whatever behavior the tool leaves behind, not the behavior the original lab shipped and tested.