MiniMax open-sources H3 video model, but 2K stays behind its API

MiniMax just open-sourced the weights for H3, its latest video-generation model. The release includes two H3-Base variants: one for text-to-video and first/last-frame generation, and one for more complex tasks that reference specific people, actions, video, and audio. You pick whichever fits the job — no need to run both.

Under the hood, H3-Base packs a 33-billion-parameter dense single-stream Transformer as its generation backbone, with an encoder based on Qwen3-VL-32B. Run it locally and you can generate 4- to 15-second clips at 768p, complete with stereo sound.

But this isn't the full H3 system. MiniMax is holding back H3-Context-IR, which understands relationships between complex inputs; H3-Regenerate-2K, which re-renders 768p output into 2K; and the sparse attention implementation. To reproduce the official 2K quality, developers will still need to call the MiniMax API.

There's also the small matter of licensing. The MiniMax H3 Community License doesn't cover the US, EU, UK, or South Korea — those regions require a separate arrangement. And if your company pulls in more than $20 million in annual revenue, you'll need MiniMax's written authorization up front.