Heretic

Heretic is an open-source tool for training-free modification of language models, with more than 37,000 stars on GitHub and an AGPL-3.0 license. Instead of merely running a model, it lets you change how a model behaves, using mathematical steering techniques rather than expensive retraining, on hardware as modest as a gaming PC.

Overview

Heretic works on any language model you can run on your own hardware. It steers behavior from contrasting example prompts and matching scorers: a hyperparameter optimizer built on Optuna searches for settings that move the chosen scores in the direction you want, and a pluggable modifier applies them to the weights. No training run is involved. On an RTX 3090, processing a 4-billion-parameter model with the default configuration takes about 30 to 40 minutes. After processing, you can save the model, upload it to Hugging Face, or run standard benchmarks, and the community has already published thousands of Heretic-processed models.

Key Features

• No training required: modify a model in an afternoon on a single GPU, instead of the compute budget only large labs can afford.
• Ready-made behavior recipes: the default configuration suppresses refusals, while bundled configurations reduce repetitive slop, humor, and sycophancy, or shift emotional alignment.
• Two built-in modifiers: Arbitrary-Rank Ablation, developed specifically for Heretic, and classic directional ablation, with flexible ablation-weight kernels and norm preservation.
• Pluggable scorers: behavior is driven by scores such as refusal rate and KL divergence, and you can write your own scorer plugins to steer any trait you can measure.
• Broad model support: most dense models, several mixture-of-experts architectures, many multimodal models, and modern hybrid models are supported.
• Reproducible results: processed models can carry a small manifest that lets anyone recreate the identical model on demand, forming the community Heretic Grimoire library of reproducible models.
• Benchmarked, not just claimed: Heretic models have been independently evaluated on standard benchmarks such as MMLU and GSM8K and compare favorably with models from competing tools.

💳 Pricing

Heretic is completely free and open source under the AGPL-3.0 license. You install it with pip, run it locally, and there is no paid tier or account.

Who Is It For

Heretic is for researchers, developers, and model enthusiasts who run open models locally and want to study or adjust their behavior. It suits anyone curious about model steering, teams building custom variants without a training pipeline, and hobbyists who want a model tuned to their own preferences.