H3-Tiny for ComfyUI - optimized MiniMax H3 weights
2 min readupdated
A package of H3-Tiny weight files packaged as single files for native MiniMax H3 support in ComfyUI. The model is the result of structural pruning (removing less important weights to speed up operation) of the base MiniMax H3 architecture for joint video and audio generation. Because ComfyUI automatically reads layer depth and width from tensors, installing third-party custom nodes is not required. This release will be useful for videomakers and researchers running video batch generation locally via ComfyUI.
Key features
- ComfyUI format: ready-to-use weight files ranging from 28.1 to 34.2 GB compatible with native MiniMax H3.
- Reduced size: models with 13.98B and 17.05B parameters require fewer resources than the original versions.
- Audio and video: support for joint generation of video sequences and soundtracks.
- No custom nodes: architecture parameters are read automatically from checkpoint files.
What can you do
- Image-to-video generation: use
fl2vafiles to create video from an image (T2VA / FL2VA). - Working with references: apply
ref2vafiles for Reference to Video scenarios in ComfyUI. - Local execution: deploy the video generation model on modern hardware with bf16 support.
Example
Place the model file in the ComfyUI/models/diffusion_models/ folder, add text encoders and VAE, then select the model in the Load Diffusion Model node and run the sampler (50 steps, simple scheduler, and euler are recommended).
Read more
DetailsAvailable options: 17.05B (Ref2VA and FL2VA / T2VA) and 13.98B (OKD-Ref2VA).
Requirements: tested on a single H200 accelerator at 1344×768 resolution and 243 frames.
License: distributed under the terms of the MiniMax H3 Community License Agreement.














Comments