MiniMax H3 Generates 2K Video With Native Stereo Audio in One Shot

MiniMax launched H3 on July 31 — also distributed as Hailuo H3 or Hailuo 3.0 — and the meaningful part isn't resolution or clip length. It's that the model generates synchronised stereo audio alongside the video in a single inference pass, without a separate audio model or a post-processing pipeline.

Previous MiniMax / Hailuo models were capable video generators but specialised tools. H3 is built differently: it takes text, images, video, and audio as simultaneous inputs and generates across those modalities in one context. In practice, you can give it a reference image, a script, and an audio sample and ask for a scene that incorporates all three — rather than generating video first and audio-matching it separately.

The output spec: 2560 × 1440 (2K) at 24fps, clips from 5 to 15 seconds. MiniMax reports 2K generation pricing significantly lower than "mainstream industry alternatives" — specific numbers haven't been published in the launch materials, but the API pricing is live and listed on the MiniMax Open Platform.

Omni-referencing is the other notable feature: the model can take multiple reference images, videos, and audio tracks simultaneously and use them to anchor style, character, or environment across a generated clip. Combined with instruction-based editing — swapping a character or background via natural language without re-generating the whole scene — it positions H3 toward commercial creative workflows: advertising, e-commerce product visualisation, and film pre-vis.

Access is available now via the Hailuo AI app and the MiniMax Open Platform API. Open weights are committed for release "in the coming days" following the July 31 launch, though no exact date has been confirmed. MiniMax has not disclosed the model's parameter count or architecture in the launch materials.

The model should not be confused with Kling O3, a separate video generation model from Kuaishou released in the same period.