MiniMax posts H3, a 33B omni-modal video model, to Hugging Face

A Chinese lab quietly reframing the video-generation stack is more interesting than another leaderboard chart, and that is what MiniMax has done with H3. The model card on Hugging Face describes H3 as a "general-purpose, omni-modal generative system" — one transformer that takes text, images, video, and audio as input and produces video with native stereo sound as output, rather than stitching a separate T2V, I2V, and TTS model together.

The technical shape is worth reading before the marketing. The core is a 33B-parameter dense single-stream Transformer called H3-Omni-Transformer, of which about 20B parameters are active at inference (13B AdaLN parameters are precomputed and cached). It borrows the full pretrained weights of Qwen3-VL-32B as its encoder, uses 3D multimodal rotary position embeddings for time and space, and runs in BF16. Output is 4 to 15 seconds at 24 FPS, up to 2K via a second module called H3-Regenerate-2K, with 32 kHz stereo audio from a dedicated H3-AudioVAE. Two open variants ship: a first-and-last-frame model and an "omni-reference" one that accepts up to nine images, three video clips, and three audio clips.

The honest caveat is how much is actually open. Of the three declared modules, H3-Context-IR (the multimodal instruction-refinement layer) is "hosted, not open-sourced", and H3-Regenerate-2K, the piece that lifts output to 2K, is "not yet open-sourced". MarkTechPost and other outlets report the release under a MiniMax Community License that permits commercial use for organizations under about $20M in revenue, with weights staged in the days after the July 31 announcement, so treat the licensing specifics as reported rather than settled. The card also flags that native sparse attention, mentioned in the architecture, is not in this initial release.

What the reporting does not give you is a clean apples-to-apples against Veo, Sora, or Kling, or a hardware bill for running the 33B model at 2K. But the shape of the release is the part worth watching: day-one recipes for vLLM, SGLang, Diffusers, and ComfyUI, an omni-modal input surface, and audio that comes out of the same model rather than a bolt-on. For anyone building short-form video tooling on open weights, that is a meaningful consolidation, even before the 2K piece opens.