Quantized INT4 ConvRot (W4A4) text encoder weights for MiniMax-H3-Image in ComfyUI. Built for the smallest single-file size with native comfy_kitchen support (no GGUF).

Files:

- qwen3vl_32b_minimax_h3_int4_convrot.safetensors (13.30 GB)

Main 50-layer text encoder. 51% smaller than int8_convrot (27.14 GB) and smaller than nvfp4_awq (15.69 GB).

- qwen3.5_9b_int4_convrot.safetensors (5.42 GB)

Optional Qwen3.5-9B prompt enhancer for the Generate Text node.

Setup:

1. Drop the .safetensors file(s) into ComfyUI/models/text_encoders/

2. Main encoder: load qwen3vl_32b_minimax_h3_int4_convrot.safetensors in CLIPLoader (type: minimax) and connect to MiniMaxH3ImageConditioning.

3. Optional prompt enhancer: load qwen3.5_9b_int4_convrot.safetensors in CLIPLoader (type: minimax) and connect to Generate Text (TextGenerateV2).

Links:

- Hugging Face: https://huggingface.co/Merserk/MiniMax-H3-Image-Text-Encoder-INT4-ConvRot

- Base model: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image

- ComfyUI weights & workflows: https://huggingface.co/Comfy-Org/MiniMax-H3-Image

License:

MiniMax H3 Image Community License Agreement: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/blob/main/LICENSE