Heretic
Open-source tool that automatically removes censorship from transformer-based language models using directional ablation.
What is Heretic?
Heretic is an open-source tool that removes censorship (safety alignment) from transformer-based language models without expensive post-training. It combines directional ablation (abliteration) with a TPE-based parameter optimizer powered by Optuna. The tool works completely automatically—finding high-quality abliteration parameters by co-minimizing refusals and KL divergence from the original model. This produces a decensored model that retains as much of the original model's intelligence as possible. Users report that Heretic versions achieve the same level of refusal suppression as manual expert abliterations but with significantly lower KL divergence, indicating less damage to model capabilities.
Best for: Automatic model decensoring · Research on model safety · Local LLM experimentation
Developer Stack Fit
Quick read on where Heretic fits in a software team's AI stack. Validate final fit against your codebase, data policy, and deployment model.
- Stack layer
- Secondary/general AI
- Deployment model
- Open-source deployable
- Open-source status
- Yes or source-available
- API support
- Not a primary API tool
- MCP support
- No MCP signal found
- Security posture
- Review vendor privacy and data retention
- Best use case
- Automatic model decensoring
Key Features
- 01
Fully automatic censorship removal from language models
Automatic decensoring—no expertise required
- 02
Directional ablation (abliteration) implementation
Retains more original model capability than manual approaches
- 03
TPE-based parameter optimization via Optuna
Works with many dense, multimodal, and MoE models
- 04
Co-minimizes refusals and KL divergence
A core development capability that teams use daily.
- 05
Automatic hardware benchmarking for optimal batch size
A core development capability that teams use daily.
- 06
Model quantization support (bnb_4bit)
A core development capability that teams use daily.
- 07
Built-in model evaluation
A core development capability that teams use daily.
- 08
Upload to Hugging Face directly
A core development capability that teams use daily.
- 09
Chat with decensored model to test
A core development capability that teams use daily.
Pros & Cons
What stands out
- Fully automatic—anyone can use it
- Achieves better KL divergence than manual expert abliterations
- Well-received by the local LLM community
- Active community has created 1,000+ Heretic models on Hugging Face
Watch outs
- Requires GPU with sufficient VRAM (12+ GB recommended)
- Decensoring takes ~45 minutes on RTX 3090 for 8B models
- Does not support SSMs/hybrid models yet
- Some novelty in model architectures not yet supported
Pricing Plans
Heretic Pricing
Choose the perfect plan for your needs. All plans include our core features with different usage limits and advanced capabilities.
Open Source
Need a Custom Solution?
Looking for enterprise features or custom pricing? Contact Heretic directly for tailored solutions.
Contact SalesMost teams land on the Open Source plan.
Alternatives
FAQ
What is Heretic and how does it work?
Heretic is a development tool that open-source tool that automatically removes censorship from transformer-based language models using directional ablation.. It uses AI to help users improve productivity through analyzing input and generating relevant output.
Is Heretic free to use?
Heretic offers a completely free plan. You can get started without paying anything.
Is there a free plan or trial?
Heretic doesn't offer a traditional free trial, but provides a money-back guarantee on paid plans.
What can Heretic do?
More development Tools
Cursor
AI-powered code editor with autonomous agents, multi-model support, and Automations for triggering agents via code changes, Slack, or timers.
Read review →
TurboQuant
Revolutionary KV cache compression achieving 6x memory reduction and 8x speedup for LLM inference with zero accuracy loss.
Read review →
Ollama
Local-first LLM runtime for running models on your hardware with local privacy, no per-token API costs, and offline-capable workflows.
Read review →Affiliate Disclosure: We may earn a commission when you purchase through links on our site. This doesn't affect our editorial independence or the price you pay.
Heretic
Free