Using minimax text encoder forTextGenerate node
I've tried to use the qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors text encoder with the text-generate node to use as a prompt enhancer, but I only get gibberish (see image). I've set the clip type to minimax. I also tried text-only generation with the same result. Is this supported?
It is not supported at this point, the minimax version doesn't have all the layers or the lm_head to get the size down.
Ah, ok! With these big models is it probably a better idea to have a seperate workflow for prompt enhancement anyway. Thanks for the info!
i think you can use these (they are already uncensored too):
https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot/tree/main
the generation tail has the layers required and should work with the builtin textgenerate node.
the other one is used for text conditioning on the h3 model. havn't tried myself though, still downloading.
edit: i stand corrected, doesn't work with bare comfyui, you need this node apparently: https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide