{"name":"heretic","repo":"p-e-w/heretic","url":"https://midorreal.com/project/p-e-w-heretic","repository":"https://github.com/p-e-w/heretic","description":"Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more","facts":{"license":{"value":"AGPL-3.0","source":"https://github.com/p-e-w/heretic","checked":"22 Nov 2026"},"language":{"value":"Python","source":"https://github.com/p-e-w/heretic","checked":"22 Nov 2026"},"homepage":{"value":"https://heretic-project.org","source":"https://heretic-project.org","checked":"22 Nov 2026"},"last_release":{"value":"v2.1.0, 8 Nov 2026","source":"https://github.com/p-e-w/heretic/releases","checked":"22 Nov 2026"},"releases":{"value":"7","source":"https://github.com/p-e-w/heretic/releases","checked":"20 Nov 2026"},"contributors":{"value":"43","source":"https://github.com/p-e-w/heretic/graphs/contributors","checked":"20 Nov 2026"},"archived":{"value":"no","source":"https://github.com/p-e-w/heretic","checked":"22 Nov 2026"}},"stars":42533.0,"writeup":{"what_it_is":"Heretic is a tool that removes safety alignment restrictions from transformer-based language models using directional ablation and automatic parameter optimization. It modifies model weights to reduce refusals while preserving the model's original capabilities.","audience":"Researchers and users working with language models locally.","claims":[{"kind":"broad","claim":"Works completely automatically without requiring understanding of transformer internals","excerpt":"This approach enables Heretic to work **completely automatically.** Heretic finds high-quality abliteration parameters by co-minimizing the number of refusals and the KL divergence from the original model.","status":"too_broad"},{"kind":"specific","claim":"Supports most dense models including multimodal and MoE architectures","excerpt":"Heretic supports most dense models, including many multimodal models, several different MoE architectures, and even some hybrid models like Qwen3.5.","status":"not_checked"},{"kind":"specific","claim":"Supports model quantization with bitsandbytes to reduce VRAM requirements","excerpt":"Heretic supports model quantization with bitsandbytes, which can drastically reduce the amount of VRAM required to process models. Set the `quantization` option to `bnb_4bit` to enable quantization.","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-10-09"},"why_now":null,"owner_supplied":null,"score":{"verdict":"too_early","hype":98,"reality":66,"gap":32,"momentum":1.5,"confidence":85,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":{"estimate":true,"lean":"too_close","gap_low":24,"gap_high":46,"rule":"early-2.0"},"changes":[{"date":"2026-11-08T07:14:03+00:00","text":"Latest release v2.1.0","source":"https://github.com/p-e-w/heretic/releases"},{"date":"2025-11-16T13:00:29+00:00","text":"First release","source":"https://github.com/p-e-w/heretic/releases"},{"date":"2025-09-21T05:46:16+00:00","text":"Repository created","source":"https://github.com/p-e-w/heretic"}],"cite":"heretic is a GitHub project at https://github.com/p-e-w/heretic written mainly in Python, described there as \"Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more\".","attribution":"Data from Mid or Real, https://midorreal.com/project/p-e-w-heretic, checked 22 November 2026. Attribution with a link is required.","still_real":null}