MiniMax-H3: 15 million downloads: MiniMax H3, ready to plug into ComfyUI

This package provides the necessary files to run the MiniMax H3 video generation model in ComfyUI. Separate from the original model repository, this package is structured to allow placement of components ranging from the text encoder to the diffusion model and VAE according to the ComfyUI folder structure. It includes a text encoder based on Qwen3-VL-32B and offers quantization options that are not dependent on specific GPU architectures.

It supports Text-to-Video (T2V), Image-to-Video (I2V), and Reference-to-Video (R2V) generation workflows. Designed to be compatible with official Comfy-Org workflow templates, it can be used immediately without additional configuration. Model files are available in various precision and quantization versions, such as bf16, int8_convrot, and fp8_scaled, providing a wide range of choices based on hardware specifications.

What distinguishes this from existing approaches is that it has been repackaged specifically for ComfyUI. While the original MiniMax H3 model may require separate environment setup, this repository is optimized so that files are recognized immediately when placed in the sub-structure of ComfyUI's models folder. In particular, the nvfp4_awq quantized text encoder can be used in environments other than Blackwell GPUs, improving accessibility.

This is suitable for developers who use ComfyUI as their primary video generation tool. It is useful for quickly testing the MiniMax H3 model in a local environment or for building custom workflows by combining it with various LoRAs. With over 15 million downloads, it is a model actively used by the community.