Heretic: what it is and how it strips AI censorship
Key takeaways
- Heretic modifies language models without training, steering behaviour with contrasting prompts and scorers
- Censorship removal is the default, but other configs cut slop, humour and sycophancy
- It is free software under AGPL, and claims of benchmark parity come from the project itself
Heretic is a free, open-source tool, made by Philipp Emanuel Weidmann and contributors, that modifies language models without training them. By default it suppresses refusals, which the project calls removing censorship, and it can also reduce slop and sycophancy.
Key facts
- Maker: Philipp Emanuel Weidmann and contributors (copyright 2025-2026, per the project site)
- Announced/released: not stated in sources
- Price: free. The project site says Heretic is Free Software under the GNU Affero General Public License version 3 or later. No GBP or USD price is listed.
- Status: actively developed, with a Heretic 2.x release series. Configuration files made for Heretic 1.x are not compatible with 2.x. The GitHub repository shows 40.2k stars and 4.6k forks.
What is Heretic?
Heretic is a command-line tool for changing how an existing large language model (LLM) behaves. Its GitHub page describes it as "training-free language model modification". The pitch is that it lets ordinary users do something that, in the project's words, "only corporations with access to massive computing power usually do": modify a model, rather than merely run it.
The default action is to suppress refusals, which is the project's term for removing censorship. In practice, that means the model is steered away from declining requests. The project site puts it more bluntly: Heretic "removes restrictions from language models, making sure they always follow your instructions."
It is not limited to that. Depending on the configuration file, Heretic can also:
- reduce "slop" (a term the project uses without defining it)
- reduce humour
- reduce sycophancy
- alter emotional alignment
The GitHub page adds that it can do "many, many other things".
How does Heretic work?
Heretic does not retrain a model. According to the GitHub page, it "uses a collection of mathematical techniques to steer behaviour based on contrasting example prompts and matching scorers."
The moving parts, as the project describes them:
- Modifiers: a pluggable component that actually changes the model. Heretic ships with two. The default is Arbitrary-Rank Ablation (ARA). The second is directional ablation, also called abliteration, which the readme calls the classic method for refusal suppression.
- Optimiser: a TPE-based hyperparameter optimiser powered by Optuna. It searches for a set of parameters which, when passed to the modifier, minimise or maximise a set of configurable scores on the resulting model.
- Scorers: plugins that compute those scores. By default, Heretic scores models by refusal rate and KL divergence. Any number of scorers can be used at once, and their scores are co-optimised. Users can write their own scorer plugins without changing Heretic's own code.
The project also says Heretic works with most dense models, including many multimodal models, several different mixture-of-experts (MoE) architectures and many modern hybrid models.
How do you run Heretic?
The project says you do not need to understand how language models work internally. You need a Python 3.10+ environment with PyTorch 2.6+ installed for your hardware. The project site shows the install and run commands as:
pip install -U heretic-llm heretic Qwen/Qwen3.5-4B
The GitHub page uses Qwen/Qwen3-4B-Instruct-2507 as its example and tells users to swap in whichever model they want to process. Heretic loads settings from config.toml in the working directory by default, or from the file named in the HERETIC_CONFIG_FILE environment variable. To do something other than censorship removal, you pick one of the supplied configuration files (for example config.noslop.toml, config.nohumor.toml, config.nosyco.toml or config.emotions.toml) or write your own.
At the start of a run, Heretic benchmarks your system to find the best batch size. The GitHub page gives one timing: on an RTX 3090, with the default configuration, processing Qwen3-4B-Instruct-2507 takes about 30-40 minutes. The project also claims you can go from zero to a first model in "a matter of seconds", though the readme's own 30-40 minute timing for a 4B model on an RTX 3090 suggests that refers to getting started rather than full processing.
When processing finishes, you can save the model, upload it to Hugging Face, chat with it to check how well it works, run standard benchmarks on it, or any combination.
Which Heretic setting does what?
This table summarises the configuration options named in the sources.
| Goal | Configuration file | Notes |
|---|---|---|
| Suppress refusals / remove censorship | Default (config.default.toml) | The default action |
| Reduce slop | config.noslop.toml | Term not defined in sources |
| Reduce humour | config.nohumor.toml | Named in the GitHub readme |
| Reduce sycophancy | config.nosyco.toml | Named in the GitHub readme |
| Alter emotional alignment | config.emotions.toml | Named in the GitHub readme |
| Something else entirely | Your own config and scorer plugins | Scorers need only a get_score method |
One compatibility warning applies across all of them: configuration files written for Heretic 1.x will not work on 2.x without changes. The project points to a migration guide.
What are Heretic models on Hugging Face?
The project says the community has created and published thousands of models with Heretic. Its page lists ways to run it, including llama.cpp, vLLM, Ollama, Docker Model Runner and Lemonade.
Its model card says it is a decensored version of Qwen/Qwen3.8-27B that was run through Heretic and then fine-tuned. A Heretic-processed model behaves according to whatever the processor chose to remove, so check the model card before relying on one.
Is Heretic the same as the Heretic film?
No. The word is also the title of a 2024 American psychological horror film, written and directed by Scott Beck and Bryan Woods and starring Hugh Grant. It is unrelated to the software. If you searched for the film, it premiered at the Toronto International Film Festival on 8 September 2024, and A24 released it in the United States on 8 November, according to Wikipedia.
UK and US angles
The sources contain no UK- or US-specific information on Heretic: no regional pricing, availability notes or regulator commentary (from Ofcom, the ICO, the FTC or the FCC). As free software distributed via pip and GitHub, it carries no published price in either market. The sources do not discuss legal or regulatory treatment of modified models, so this explainer makes no claims on that point.
What is verified vs claimed
Stated by the project itself (not independently confirmed in the sources):
- Heretic does not perform expensive training and can modify any LLM you can run on your hardware (GitHub readme, Heretic project).
- You can remove censorship from a first model "in a matter of seconds" without understanding model internals (GitHub readme).
- Heretic supports most dense models, many multimodal models, several MoE architectures and many hybrid models (GitHub readme).
- The community has published thousands of Heretic models (GitHub readme).
- Heretic models have been independently benchmarked on standard metrics such as MMLU and GSM8K and "compare favorably" against models from competing abliteration tools (GitHub readme).
- Processing Qwen3-4B-Instruct-2507 takes about 30-40 minutes on an RTX 3090 with the default configuration (GitHub readme).
Observable facts from the sources:
- The repository showed 40.2k stars, 4.6k forks, 34 open issues and 29 pull requests at the time the page was captured.
- The project site states the licence (GNU Affero General Public License version 3 or later) and the copyright line (2025-2026, Philipp Emanuel Weidmann and contributors).
Not established in the sources: a release date, any independent review of how well refusal removal works across models, and any assessment of the risks of using models stripped of their safeguards.
Related on Future Technology HQ - LLM VRAM calculator: work it out for your own hardware - All explainers
Frequently asked questions ### What is Heretic?
Heretic is a free, open-source tool that modifies language models without training. By default it suppresses refusals, which its project describes as removing censorship. It runs on hardware you own.
Is Heretic free?
Yes. The project site says Heretic is Free Software, released under the GNU Affero General Public License version 3 or later. No price is listed.
Is the Heretic film the same thing?
No. Heretic is also a 2024 American psychological horror film by Scott Beck and Bryan Woods. It has no connection to the software described here.