Last updated: 11/19/2026Last verified: 2026-11-05
What is Heretic?
Heretic is a free, open-source tool for modifying language model behavior on your own GPU without retraining. It uses scorer-guided ablation optimization to target behaviors such as refusals, slop, sycophancy, humor, and emotional alignment. Its workflow combines configuration presets, contrast prompts, optimization metrics, and tools for testing and exporting modified models.
Key Features
- Training-free behavior steering: Apply configuration presets that target refusals, slop, sycophancy, humor, or emotional alignment without running a retraining process.
- Scorer-guided ablation optimization: Evaluate candidate modifications over contrast prompts using metrics such as refusal rate and KL divergence.
- Pareto-front trial selection: Use a trial picker to explore trade-offs between optimization objectives rather than relying on one score alone.
- Consumer GPU operation: The supplied specifications estimate approximately 2.5 GB of VRAM per billion model parameters, with optional 4-bit bitsandbytes quantization.
- Model evaluation and export: Save, benchmark, chat-test, or upload modified models.
- Reproducible manifests: Use Heretic Grimoire manifests to document modifications and support reproducible experimentation.
Best For
Pricing
Heretic is listed as free and open source in the supplied information. Running it requires access to suitable GPU hardware; free software does not necessarily mean cost-free execution. No paid plans or hosted-service prices are provided.
Pros & Cons
Pros
- Modifies targeted model behaviors without requiring retraining.
- Provides presets for several behavior targets rather than focusing only on refusals.
- Combines optimization metrics with Pareto-front trial selection to help examine trade-offs.
- Includes benchmarking and chat-testing options alongside model export.
- Supports reproducibility through Heretic Grimoire manifests.
- Offers optional 4-bit quantization for experiments with constrained GPU memory.
Cons
- Requires suitable GPU access and technical familiarity with language model experimentation.
- Memory requirements increase with model size, which can limit experiments on smaller GPUs.
- Behavior-targeting presets do not establish that every unwanted response pattern will be removed.
- Optimizing refusal rate or KL divergence does not, by itself, establish overall model quality or safety.
- The supplied information does not identify supported model architectures or provide independent performance benchmarks.
Alternatives
An open-source mechanistic interpretability library for inspecting transformer internals and conducting activation-level experiments. It is relevant for researchers investigating behavior changes, although it is not a direct replacement for Heretic's preset-driven optimization workflow.
An open-source library for interventions on neural model representations. It suits researchers who want to design custom representation-level experiments rather than use a dedicated behavior-steering workflow.
A library for parameter-efficient model adaptation, including LoRA-based workflows. It is an alternative for users willing to train adapter parameters instead of modifying behavior through training-free ablation.
FAQ
Details
Platform
Features
- Training-free behavior steering: suppress refusals, slop, sycophancy, humor, or emotional alignment via configuration presets
- Scorer-guided optimization (refusal rate, KL divergence) over contrast prompts with a Pareto-front trial picker
- Runs on consumer GPUs with about 2.5 GB VRAM per billion parameters plus optional 4-bit bitsandbytes quantization
- Save, benchmark, chat-test, or upload modified models with reproducible manifests (Heretic Grimoire)
Languages
Known limitations
- Requires Python 3.10+, PyTorch 2.6+, and a GPU with roughly 2.5 GB of VRAM per billion parameters; CPU-only mode is orders of magnitude slower
- Distributed under AGPL-3.0, and processed models remain subject to the upstream model license and terms
- Optimization runs about 100 trials — the tutorial example took around 1.5-2 hours on a mid-range GPU per model
- Documentation and community resources are English-only








