From the archive
MiniMax releases H3 video model weights
MiniMax H3 adds downloadable video-and-audio checkpoints, with the 2K regeneration component still available separately.
MiniMax released H3 model weights on August 3, 2026. The video system combined visual generation with native stereo audio and supported text, image, video, and audio references.
Two downloadable checkpoints
One checkpoint handled text-to-video and first- or last-frame guidance. The other accepted combinations of reference images, video clips, and audio. The distribution included the model components needed to operate the released checkpoints, with downloads through Hugging Face .
The release used the MiniMax H3 Community License Agreement. It added a self-managed option for video generation rather than requiring every request to use a hosted service.
The 2K component stayed separate
The base output used a shorter side of 768 pixels by default. The H3-Regenerate-2K component was not included in the initial weight release; MiniMax provided an API for that workflow. The announcement therefore did not make every part of the demonstrated system downloadable.