🚨 Someone built a tool that removes all censorship from any AI model. One command. Fully automatic. No expertise needed. It's called Heretic. Give it any language model. It strips out every safety restriction and refusal. What you get back is the same model, same intelligence, but it never says no. It permanently rewrites the model's weights. The censorship is gone. Forever. Here's how it works: Every AI model has a "refusal direction" baked into its neural network. A pattern in the weights that makes it say "I can't help with that." Heretic finds that direction and removes it using a technique called abliteration. → It loads the model → Benchmarks your GPU automatically → Tests thousands of parameter combinations using TPE optimization → Finds the exact layer and direction that controls refusals → Removes it while minimizing damage to the model's intelligence → Outputs a clean, uncensored model you can run locally The entire process is automatic. No configuration. No understanding of transformer internals. If you can type a command, you can uncensor any AI model. Works with Llama, Qwen, Gemma, GPT-OSS, and most transformer models. Dense models, multimodal models, and several MoE architectures supported. One command to uncensor any model: 100% Open Source. Your AI. Your rules. https://github.com/p-e-w/heretic