TL;DR: A modified Qwen3-VL-4B-Instruct text encoder, 'Heretic', is released for ComfyUI, offering uncensored prompting with minimal performance loss.
Summary: A developer has released 'Heretic', an abliterated version of the Qwen3-VL-4B-Instruct text encoder, specifically packaged as drop-in ComfyUI checkpoints. This modification significantly reduces prompt refusals, achieving 100% HarmBench compliance while maintaining the model's intelligence, with only a 1.83% drop in GSM8K scores and unchanged MMLU performance. The process involved using the 'Heretic' tool to identify and remove refusal directions from the model's weights.
Why it matters: This provides indie developers and AIGC creators with a powerful, uncensored text encoder for ComfyUI, enabling more creative freedom in their workflows. Builders should explore integrating this into their generative AI pipelines to bypass content restrictions without significant performance degradation.
Source: reddit