MiniMax H3
The first open-weights omni-modal video model — unless you live in the US, EU, UK, or Korea
What it is
A 33B dense H3-Omni-Transformer that takes text, images, video, and audio in and generates synchronized video with stereo audio out: 768p to 2K, 4–15 seconds at 24 FPS with 32 kHz sound, in FL2VA (first/last-frame) and Ref2VA (multimodal reference) checkpoints. Nothing else with weights on the table does omni-modal in, video-plus-audio out — and the ~4 million downloads a month say the ecosystem noticed, with ComfyUI integration circulating since the August 3 drop.
Why it's interesting
This is the sharpest license lesson of the season, which is why it's here despite headlines calling it 'open source.' The 17.6 kB MiniMax H3 Community License grants royalty-free commercial use — but only within an 'Applicable Territory' that expressly excludes the European Union, the United Kingdom, the Republic of Korea, and the United States. If you're reading this from any of those, the license grants you nothing. Around that hole sit an acceptable-use policy, a no-competing-models clause, a $20M revenue authorization trigger, and mandatory 'MiniMax H3' UI branding.
Use cases
- Licensed-territory production: synchronized video-with-audio generation, self-hosted
- A case study in territorial licensing — the fine print that headline coverage skipped
- Watching where 'open omni-modal video' goes next, because the capability is real
Who it's for
Builders outside the excluded territories; everyone else, counsel first — this shelf exists for you
Setup
Advanced. ~42.5GB of BF16 weights; reference deployments use 4-GPU SGLang setups, with ComfyUI workflows for the checkpoints
Limitations & cautions
The license simply does not apply in the US, EU, UK, or Korea without individually requested written authorization; clip length tops out at 15 seconds; and the no-compete and AUP obligations travel with every redistribution, NOTICE files and all.
Editorial takeaway
Four million downloads a month and four excluded jurisdictions. 'Open weights' has never needed the follow-up question more: open where? Read the territory clause before you read the benchmarks.