Open-source Heretic lets a gaming PC rewrite LLMs: about 30 to 40 minutes for Qwen3-4B on an RTX 3090
Synopsis
The open-source tool Heretic brings behavior modification of language models from large-compute organizations to personal GPUs: instead of expensive training, it uses a set of mathematical techniques to steer behavior from contrasting example prompts and matching scorers, with refusal suppression as the default. On an RTX 3090 with the default configuration, processing Qwen3-4B-Instruct-2507 takes about 30 to 40 minutes, and the community has published thousands of Heretic-produced models, with independent benchmarks comparing them against outputs of competing tools.
Interpretation
Users on personal hardware can now modify, not merely run, a language model; adjustable behaviors include suppressing refusals, reducing slop, reducing humor, reducing sycophancy and altering emotional alignment. Such modification previously belonged to organizations with massive computing power; Heretic explicitly avoids expensive training and steers behavior with mathematical techniques. On an RTX 3090 with the default configuration, processing Qwen3-4B-Instruct-2507 takes about 30 to 40 minutes, and going from zero to removing censorship from a first model starts in seconds.
The tool steers behavior by optimizing one or more scores computed by pluggable scorer plugins, and any number of scorers can run simultaneously and be co-optimized. The default scorers are refusal rate and KL divergence, the former measuring the share of refusals and the latter measuring how far the modified model's distribution departs from the original. Scorers are enabled in the main configuration file with a minimize or maximize setting, the same scorer can be configured in multiple instances with different settings, and a plugin path can be a fully qualified import path or a file path plus class name.
Heretic supports most dense models, including many multimodal models, several MoE architectures and many modern hybrid models, and after processing the user can save the model, upload it to Hugging Face, chat with it or run standard benchmarks. The community has created and published thousands of models with Heretic. Those models have been independently benchmarked with standard metrics such as MMLU and GSM8K and found to compare favorably against models produced with competing abliteration tools.
Perspective
The demonstrated scope is behavior modification of locally runnable models on consumer GPUs with the default or bundled configurations; the concrete example given is processing Qwen3-4B-Instruct-2507 on an RTX 3090 in about 30 to 40 minutes. The intended reader is a user who can set up a Python 3.10+ environment with PyTorch 2.6+.
The material states that Heretic models were independently benchmarked with MMLU and GSM8K and compare favorably against models from competing abliteration tools, but it lists no scores; configuration files made for Heretic 1.x are not compatible with 2.x and require changes per the migration guide.