MiniMax H3

MiniMax H3 is the Chinese lab's next-generation video model, released on July 31, 2026. It generates clips at native 2K up to roughly 15 seconds, and the headline capability is that it takes text, image, video and audio as inputs to the same prompt rather than making you pick a mode. You can hand it a reference image for a character, a reference clip for motion or framing, and a text instruction for what should happen — the model resolves all of them together. MiniMax calls this omni-reference; in practice it's the feature that separates H3 from the text-to-video-plus-optional-first-frame pattern most competitors still use.

The second thing that matters is audio. H3 produces sound synchronized to the generated footage in the same pass, which puts it in the small group of models — Veo 3 being the obvious comparison — that don't leave you compositing a soundtrack afterwards. Native 2K also means you aren't upscaling 720p output before it's usable anywhere near a real timeline. It's already available through third-party inference hosts including fal.ai, so you can hit it via API without going through MiniMax's own platform.

The audience here is developers and studios building video generation into a product, not casual users looking for a web app to play with — this is a model release first. Open weights are the strategic differentiator: if you can run it yourself, you can fine-tune on your own footage, keep client material off a third-party server, and avoid per-generation API costs at volume. That's a meaningfully different proposition from Veo 3 or Kling, both of which are closed. The tradeoff is that a 2K video model with audio is not something you casually self-host, and MiniMax's tooling and documentation remain thinner than what US labs ship.