Workflows: https://huggingface.co/molbal/MiniMax-H3-GGUF/tree/main/workflows
Loader nodes: https://github.com/molbal/ComfyUI-GGUF / https://registry.comfy.org/publishers/molbal/nodes/comfyui-gguf-reboot
MiniMax H3
System Overview
MiniMax H3 is a general-purpose, omni-modal generative system. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds. Thanks to its task-generalization-oriented system design, H3 already possesses broad multimodal context understanding and generation capabilities at the pre-training stage, enabling outstanding performance in following complex multimodal instructions.
H3 supports the following input and output specifications:
Category Specification Output duration 4–15 seconds Output aspect ratio Supports a wide range of aspect ratios, including but not limited to 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 Output resolution Supports various resolution dimensions. The shorter side is set to 768 pixels by default. 2K | generation can be achieved with H3-Regenerate-2K Output frame rate 24 FPS Output audio 32 kHz stereo Supported dialogue languages Stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Additional languages are also supported to varying degrees
Description
FAQ
Comments (8)
I was actually planning to ask if you intended to release a GGUF quantized version of the Minimax-H3 model, but it turns out you’ve already done it—the efficiency is truly amazing. Also, since the model's text encoder and VAE are quite large, have you considered using the GGUF format for them as well?
Thank you! I can definitely do the text encoder too as well. For that I will not make my own conversions, but rather patch the loader nodes to load Unsloth's quants.
However I ask for a bit more patience, than usual, as this is a larger model with large encoder/vae, each download/upload/trial-and-error experiment take more time
Hi @Mercury16 - added support for Qwen3VL-32B as a text encoder
@molbal Hey there—unfortunately, as I suspected, I’m unable to run the Q8_CR model on my Mac. The issue is essentially the same as the one I encountered with INT8 models:
NotImplementedError: The operator 'aten::_int_mm' is not currently implemented for the MPS device.
The INT8 implementation calls PyTorch’s torch._int_mm, but that operator hasn't been implemented for MPS yet. It seems there isn't a better workaround for this at the moment. While running diffusion models on MPS is frustratingly slow—to the point of being practically unusable—I’d still be very happy to help you run tests on macOS in the future for experimental or exploratory purposes. I also tried out the custom node you created for GGUF conversion; converting safetensors models to Q8_CR on macOS works well and is relatively fast. It’s a real shame that the resulting Q8_CR models can't actually be used, but I’ll certainly keep following your updates.
@Mercury16 thanks for testing them! That's a shame. Let's hope Int8 support is added for Mac by ComfyUI soon. Unfortunately it is way too big of a task for me to attempt within the scope of these loader nodes :(
# ComfyUI Error Report ## Error Details - Node ID: 145 - Node Type: UnetLoaderGGUF - Exception Type: ValueError - Exception Message: ValueError: Unexpected architecture type in GGUF file: 'minimax_h3'
Doesnt work
Can you double check please, that you have the molbal/comfyui-gguf package and not city96/comfyui-gguf?
This one: https://github.com/molbal/ComfyUI-GGUF / https://registry.comfy.org/publishers/molbal/nodes/comfyui-gguf-reboot
Example workflows: https://huggingface.co/molbal/MiniMax-H3-GGUF/tree/main/workflows
Thank you! Your ComfyUI node and GGUF model work perfectly.