MiniMax Officially Launches H3 Omni-Modal Video Model, Open Weights Coming Within Days
News Summary
MiniMax has moved H3 from preview to production, officially launching its next-generation omni-modal video model roughly two weeks after first showing it under a "Coming Soon" placard at WAIC 2026 in Shanghai. The model is now live through MiniMax's API and the MiniMax Hub platform, and the company says it plans to release H3's model weights publicly within the next few days, extending the open-weight approach it took with its M3 language model.
From Preview to Launch
When MiniMax showed H3 at its WAIC booth on July 17, 2026, company representatives described it only as "coming soon," with no public release date confirmed at the time. That gap has now closed: H3 is listed as a live model on MiniMax's official platform, accessible to developers through the API or for hands-on testing through MiniMax Hub.
Capabilities Confirmed at Launch
The shipped model matches the specifications MiniMax previewed at WAIC. H3 generates native 2K video at 24 frames per second, with clip lengths from five to fifteen seconds, supporting text-to-video, image-to-video, first-and-last-frame, and multimodal reference generation modes.
Its standout feature remains "omni-reference" support, letting creators supply up to nine reference images, three reference video clips, and three reference audio clips within a single request to guide style, character consistency, motion, or voice without forcing rigid keyframes. H3 also supports multi-shot storytelling for more cinematic multi-scene sequences and instruction-based editing through natural-language commands. Dialogue, sound effects, and ambient audio are generated in the same pass as the video and synced to on-screen action, removing the need for a separate audio production step in many short-form projects.
Open Weights Coming Within Days
MiniMax says it plans to open-source H3's model weights within the next few days of launch, following the pattern it set with M3: that language model was announced June 1, 2026, with weights published to Hugging Face by June 7. As with M3, an open-weight release for H3 would make the model available for self-hosting and fine-tuning, though it remains to be seen whether MiniMax will also release full training code or inference operators, which it has not done for M3.
Context Within MiniMax's Broader Lineup
H3's launch completes a pairing MiniMax previewed together at WAIC: a strong reasoning- and coding-capable language model in M3, and an increasingly capable generative multimedia counterpart in H3. Together they position MiniMax as a broad-based competitor across both language and multimodal generation, rather than a single-product specialist, and they reinforce the open-weight strategy that Chinese AI labs have increasingly leaned on this year following Jensen Huang's public endorsement of open-weight AI in July.
What to Watch Next
With H3 now live, independent benchmarking and broader creator feedback can begin in earnest, something that wasn't possible while the model was still a WAIC-floor preview. The next milestones to watch are the actual weights release MiniMax has promised within days, along with fuller pricing and technical documentation as usage scales beyond early access.