A full-modal video model opens its weights: 15 seconds, 24fps, 32kHz stereo
MiniMax H3 generates 4 to 15 seconds at 24fps with 32kHz stereo sound and selectable aspect ratios, integrating text, image, video and audio context. Open weights, community licence.
MiniMax has open-sourced H3, a next-generation general video model described as a full-modal generation system that integrates and understands multimodal context composed of text, images, video and audio. Output runs 4 to 15 seconds at 24fps with 32kHz stereo sound, with selectable aspect ratios including 21:9, 16:9, 4:3 and 1:1.
Synchronised stereo audio generated with the video, rather than dubbed onto it afterwards, is the specification worth noting. Audio-video sync is a hard generation problem — footsteps landing on the frame the foot lands, speech matching mouth movement — and treating it as one output rather than two pipelines is the harder and more correct approach.
The aspect-ratio list reads like a product decision rather than a research one. 21:9, 16:9, 4:3 and 1:1 is a set chosen for cinema, television, archive and social formats — a model shipped for people making things, not for a benchmark.
Open weights on a video model of this capability is the more consequential fact. Video generation has been the most closed of the major modalities, largely because the compute and the liability are both high. Weights in the open means fine-tuning, local deployment, and no ability for the publisher to withdraw it later — which is the point, and also the risk everyone will argue about.
The licence is the part to read before building. MiniMax ships under community terms rather than Apache or MIT, and community terms carry conditions that only bind some users, which is not the same as none.
AIBase — Visual Large Models Receive a Major Open-Source Announcement → · Digen — 8 Best Open Source AI Video Models to Use in 2026 →