Skip to content
Sign in
  • Browse all toolsFilter by category and pricing
  • Browse categoriesWriting, code, image, audio and more
  • Compare toolsPut up to three side by side
  • Read the blogGuides and comparisons

Heretic

Open-source language model modification for refusal suppression, shorter responses, and reduced sycophancy.

Heretic is a free, open-source language model modification tool that changes model behavior without training, including refusal suppression, sycophancy reduction, and shorter responses. Heretic serves local model users, developers, and researchers, and is available under the AGPL-3.0 license through a Python package and GitHub repository.

Updated 9 Nov 20266 sections

Heretic website screenshot

What is Heretic?

Heretic is an open-source project for modifying language models without training. Heretic changes model behavior through mathematical techniques guided by contrasting example prompts and matching scorers. The project is intended for people who want to modify models they can run on their own hardware, rather than only use those models for inference.

Heretic supports most dense models, including many multimodal models, several mixture-of-experts architectures, and many modern hybrid models. The repository states that understanding language model internals is not required. Heretic nevertheless operates as a Python-based tool, with environment preparation and configuration forming part of the documented workflow.

What Heretic can do

Heretic provides supplied configurations for several behavior changes, with refusal suppression as the default action. Heretic also supports custom scoring criteria and alternative modifier implementations. These controls let users define the behavior to optimize and choose the mathematical method applied during a processing run.

  • Heretic suppresses refusals through the default configuration.
  • Heretic provides configurations to reduce slop, humor, and sycophancy.
  • Heretic can alter emotional alignment and shorten responses.
  • Heretic co-optimizes multiple scores through configurable scorer plugins.
  • Heretic includes Arbitrary-Rank Ablation and directional ablation modifiers.
  • Heretic offers model saving, Hugging Face uploads, chat testing, and standard benchmarks.

How Heretic works

Heretic requires Python 3.10+ and PyTorch 2.6+ installed for the available hardware. Users install the heretic-llm package with pip and run the heretic command with a model identifier. Heretic loads config.toml from the working directory by default, or uses a configuration path specified through the HERETIC_CONFIG_FILE environment variable.

Heretic benchmarks the system at startup to select a batch size. Heretic then combines a pluggable modifier with an Optuna-powered TPE hyperparameter optimizer to minimize or maximize configured scores. Refusal rate and KL divergence are the default scoring criteria. Custom scorer plugins can be loaded without modifying the Heretic package code.

Who Heretic is for

Heretic serves local model users seeking behavior changes and developers or researchers who want configurable model modification. Supplied configurations cover common adjustments, while plugin interfaces support custom objectives. Heretic also supports people sharing modified models: optional reproduction information can accompany Hugging Face uploads, and public manifests can be collected locally through the Heretic Grimoire.

  • Heretic lets local model users process a model their hardware can run.
  • Heretic lets developers implement custom scoring criteria.
  • Heretic lets researchers benchmark processed models.
  • Heretic lets model publishers include reproduction instructions with uploads.

Heretic pricing and access

Heretic is free, open-source software distributed under the AGPL-3.0 license. Heretic is available through the GitHub repository and the heretic-llm Python package. The documented setup uses a local Python and PyTorch environment, and the selected model must fit the capabilities of the user's hardware.

Heretic strengths and limitations

Heretic combines training-free modification, configurable scoring, and model export options in a documented local workflow. Heretic also provides optional reproduction manifests, environment checks, and exported weight hash verification. Exact reproduction remains conditional: the repository explains that identical results require matching hardware, drivers, operating system, package versions, and the rest of the software environment.

  • Heretic requires hardware capable of running the chosen model.
  • Heretic permits only a single modifier in each processing run.
  • Heretic configuration files from 1.x require migration for 2.x.
  • Heretic cannot guarantee identical reproduced weights when environments differ.

Heretic FAQ

What is Heretic?

Heretic is an open-source tool for modifying language model behavior without training. Heretic uses mathematical techniques, contrasting example prompts, and configurable scorers to make changes such as refusal suppression, reduced sycophancy, altered emotional alignment, and shorter responses.

Is Heretic free?

Heretic is free, open-source software licensed under AGPL-3.0. Heretic is available through its GitHub repository and the heretic-llm Python package, with a documented setup requiring Python, PyTorch, and suitable hardware for the selected model.

What models does Heretic support?

Heretic supports most dense models, including many multimodal models, several mixture-of-experts architectures, and many modern hybrid models. Heretic requires users to choose a model that can run on their available hardware.

How do you install and run Heretic?

Heretic is installed with pip install -U heretic-llm after preparing Python 3.10+ and PyTorch 2.6+ for the hardware. Heretic runs from the command line with a model identifier and uses refusal suppression by default unless another configuration is selected.

Can Heretic upload models to Hugging Face?

Heretic can upload processed models to Hugging Face and optionally include reproducibility information. Heretic can also save models, provide chat testing, and run standard benchmarks after processing. Exact reproduction depends on how closely the reproduction environment matches the original environment.

Can Heretic use custom scoring plugins?

Heretic supports custom scorer plugins without requiring changes to the package code. Heretic configures active scorers through the main configuration file, supports simultaneous optimization of multiple scores, and allows multiple instances of the same scorer with different settings.

From the community

No reviews yet

Be the first to review this tool. What you write helps the next person choose.

More tools like Heretic

llama.cpp

www.github.com

Open-source LLM and vision-language model inference in C/C++ for local and cloud hardware.

Free

Automodel

www.github.com

Open-source PyTorch distributed training and fine-tuning for language and vision-language models

Free

Mistral Inference

www.github.com

An archived, open-source inference library for running Mistral models through Python and command-line tools.

Free

DeepSeek Janus

www.github.com

Open-source multimodal models for image understanding and text-to-image generation

Free