MiniMax’s H3 video model went from an unanswered question to a downloadable checkpoint on 3 August. AI Insiders reported two days earlier that MiniMax had announced open weights for H3 without publishing a weights link, a parameter count, or any way to actually get the model. That gap is now closed. The weights are live, H3 carries 33 billion parameters, and Artificial Analysis has already benchmarked it against the field.
The answer is only partial. Artificial Analysis’s rankings put H3 in first place for Video Editing, second for Text-to-Video, and third for Image-to-Video, making H3 the first open-weight model to lead a category in that ranking. The system takes images, audio, video, and text together, producing clips from four to 15 seconds with stereo sound. A single prompt can pull in as many as nine reference images, plus three separate video clips and three separate audio clips.
Two components did not ship with the release, and their absence is not incidental. H3-Context-IR and the 2K resolution module both stay closed. H3-Context-IR converts a prompt plus any reference photos, video, or audio into the format the model needs before it renders, and the resolution module is what pushes output past what the open weights can reach on their own. Run H3 locally through ComfyUI and output caps at 768p. Matching the higher end of what MiniMax’s own hosted service can do means manually managing context prep with the company’s published guides instead of the withheld intermediate format.
That split matters more than the ranking for anyone deciding whether to build on H3. The open checkpoint can still be fine-tuned on a company’s own footage, a specific character, or a house visual style, which is useful for teams training a specialized version. But the license caps commercial use at $20 million in annual revenue, a ceiling that makes H3 non-free by the standard open-source definition the moment a company clears that line.
The timing sharpens the contrast with ByteDance, whose closed Seedance 2.5 shipped the same day, producing 30-second clips that include built-in audio and doubling H3’s maximum length. AI Insiders has covered that release separately. Set against it, MiniMax opening any weights at all, even partial ones, is itself a signal about how the two companies are competing for developer attention right now.
Teams evaluating open video models should treat H3’s ranking as proof of the underlying architecture, not proof of what shipping a product on it will look like. Budget for the $20 million revenue ceiling before scaling past a prototype, and plan to rebuild the context-formatting step MiniMax kept for itself.
This account draws on reporting by Maximilian Schreiner for The Decoder, published 3 August 2026.
