Qwen3-VL-8B-Heretic (INT8 / W4A8 Convrot) – Abliterated Text Encoder for Qwen Image
This is a quantized conversion of the excellent DreamFast/Qwen3-VL-8B-Heretic-1.3.0 model (an abliterated version of Qwen3-VL-8B-Instruct created with Heretic v1.3.0).
I converted it into:
INT8 Convrot
W4A8 Convrot
These formats were not available before, so I made them for easier and more efficient use as a text encoder in ComfyUI / Qwen Image workflows.
Original model source:
https://huggingface.co/DreamFast/Qwen3-VL-8B-Heretic-1.3.0
All credit goes to the original author (DreamFast). This is only a quantization conversion — I did not train or abliterate the model myself.
Usage:
Place the files in your ComfyUI/models/text_encoders/ folder and load them as usual.
Enjoy the smaller VRAM footprint!
Description
FAQ
Comments (5)
is it possible to convert Convrot to GGUF?
unfortunately, only gguf and only Q3-Q4 will work on my system😥
Upd: GGUF works after placing mmproj-Qwen3VL-8B-Instruct-Q8_0 in the models\text_encoders folder—alongside Qwen3-VL-8B-Q4_K_S.gguf—and installing the ComfyUI-GGUF fork by leejet (https://github.com/leejet/ComfyUI-GGUF).
The hugging face link provided has .gguf's
update
Thanks. I wonder why the original uploader only provided quantized GGUF versions instead of the converted INT8 files? (The release page only lists fragmented files like model-00001-of-00004.safetensors.) Also, I am curious: are QWEN3VL (which integrates the MMP vision component) and QWEN3-VL (which requires the MMP vision component to be loaded separately) interchangeable?
Does it make any difference? Any test images with one and another model? As technically, the model is not generating anything, so abliteration has no effect here unless diffusion reacts to the refusal vector itself.