NewOpen weights
MiniMax H3
A general-purpose omni-modal video generation system that accepts text, image, video, and audio context and produces video with native stereo audio.
A video model built around multimodal context
MiniMax describes H3 as a unified generative model rather than a single text-to-video endpoint. Its published workflows include text-to-video, image-guided generation, reference conditioning, video continuation, and audio-aware generation.
Evidence note
This page summarizes linked first-party documentation. ModelRun Lab has not independently reproduced the performance or quality claims.
Choose the route that matches your goal
01
ComfyUI native workflow
Official ↗02Best starting point for a visual local workflow. Follow ComfyUI’s version-specific guide.
Download model files
Official ↗03Review the model card, repository structure, license, and current file variants.
Repository and technical notes
Official ↗Use the upstream repository for current implementation details and changes.
What we can and cannot say yet
QuestionCurrent answerEvidence
Can it run locally?Yes, supported workflows exist● Official
Does it run on consumer GPUs?Community reports exist across several tiers● Community
Is 12 GB consistently sufficient?No reliable universal answer● Unknown
Exact generation speed?Depends on build, resolution, duration, and offload● Variable
Community reports are discovery evidence, not a guarantee that the same configuration will work on another system.