There are already many "heretic" versions of Qwen3 8B VL Instruct
Qwen Image 2.1 is based on the original, unmodified Qwen3-8B-VL-Instruct. There are already numerous "heretic" versions of the model listed under its fine-tune category:
I initially assumed this was intended for ComfyUI, but that turns out not to be the case; what is the point of placing it here?
If you are simply releasing a "heretic" version of Qwen3-8B-VL-Instruct, it should be listed under the fine-tunes for Qwen3-8B-VL-Instruct, not under those for Qwen Image 2.1.
Thanks — fair point, and you're right. I byte-compared the text_encoder/ weights against Qwen/Qwen3-VL-8B-Instruct (attention + MLP projections + embeddings) and they're identical, so this is stock Qwen3-VL-8B-Instruct. I've updated base_model to point at Qwen/Qwen3-VL-8B-Instruct so it shows up under that lineage.
I've kept the Qwen-Image-2.1 name because the repo is packaged specifically as a drop-in text encoder for that pipeline: the sibling repos ship it in the formats the Qwen-Image node expects (CLIPLoader type qwen_image for NVFP4 / W4A8, CLIPLoaderGGUF + mmproj for GGUF), which a generic Qwen3-VL-8B checkpoint doesn't provide. This particular repo is the plain transformers checkpoint, which is why it didn't load in ComfyUI for you. Appreciate the nudge.