September 21, 2026
Removing restrictions from an open model is now a half-hour job: Heretic does it on a single RTX 3090
The script finds and removes a model's refusal reflex on its own, while the result stays closer to the original than manual versions.
The original gemma-3-12b refused 97 requests out of 100. The same model after Heretic refused three.
Heretic removes the refusal reflex from an open model's weights. This kind of edit was done manually before, but parameters were picked by eye, and the model drifted away from the original. The community has already published more than 5000 models through Heretic.
The losses became visible. The tool measures how much an edited model has diverged from the original in behavior: the smaller the number, the closer it is. The best manual version of gemma-3-12b scores 0.45; the Heretic version scores 0.16 with the same number of refusals (author's measurement, project README, 22.09.2026).
Installed with two commands. First `pip install -U heretic-llm`, then `heretic` and the address of any Hugging Face repository. Running Qwen3-4B with the default config takes 20–30 minutes on one RTX 3090. The tool then asks what to do with the result: save it locally, upload it to Hugging Face, open a chat, or run benchmarks.
The author lowered the hardware threshold through releases. The first public version appeared on 16.11.2025; a month later, 1.1.0 added Apple Silicon, multiple GPUs at once, and support for Colab and Kaggle. Version 1.2.0 from 14.02.2026 brought 4-bit quantization, while 1.3.0 from 05.05.2026 cut peak video memory use. The current 1.4.0 was released on 14.06.2026 and requires Python 3.10+.
The tool already handles hybrids such as Qwen3.5, but state-space architectures are not supported yet.
