Heretic

Overview

Heretic is an open-source tool that removes safety alignment restrictions from transformer-based language models through a technique called directional ablation. Unlike traditional fine-tuning approaches that require expensive GPU compute and curated datasets, Heretic modifies model weights directly by orthogonalizing transformer component matrices against computed residual directions. The tool uses Optuna’s TPE sampler to automatically optimize ablation parameters, requiring no knowledge of transformer internals from the user. With over 33,000 GitHub stars, it has become one of the most widely adopted tools in the uncensored model community.

Key Capabilities

Heretic’s primary strength is its automation. Users point it at a model and it handles the rest: computing difference-of-means between harmful and harmless prompt residuals, optimizing ablation weights per layer, and co-minimizing refusals while preserving model capabilities through KL divergence tracking. The tool supports a broad range of architectures including standard dense transformers, multimodal models, Mixture-of-Experts architectures, and hybrid designs like Qwen3.5. Post-processing options include saving modified weights locally, uploading directly to Hugging Face, running interactive chat sessions for testing, and executing benchmarks to verify capability retention.

Use Cases

Researchers and developers use Heretic to create uncensored model variants for applications where safety filters interfere with legitimate use cases, such as creative writing, red-teaming and security research, academic study of model behavior, and building applications that need unrestricted model outputs. The tool is also used for studying alignment techniques by examining how ablation affects model behavior across different architectures. The automatic optimization makes it accessible for producing custom model variants without deep expertise in transformer internals.

Considerations

Heretic is released under the AGPL-3.0 license. The tool modifies model weights to remove safety guardrails, which carries significant ethical and legal implications depending on jurisdiction and use case. Users are responsible for ensuring their use complies with applicable laws and the original model’s license terms. Processing requires sufficient GPU memory to load the target model, and optimization runs can take considerable time depending on model size. The quality of results varies by model architecture and the specific safety mechanisms employed.

Who It’s For

Heretic is designed for AI researchers studying alignment and safety mechanisms, red-team security professionals who need unrestricted models for testing, developers building applications where content filters create unacceptable limitations, and the open-source community that produces uncensored model variants on Hugging Face. Users should have a clear understanding of the ethical implications and legal requirements surrounding uncensored model deployment.