MiniMax released H3, an omni-modal generative system that produces video with native stereo audio from text, image, video or audio input, according to the project's GitHub repository.
The open-sourced core, H3-Base, is a 33 billion parameter dense transformer with about 13 billion additional parameters in its conditioning branches. It uses separate visual and audio encoders tied together with three-dimensional rotary position embeddings, the repository says.
H3-Base ships in two checkpoints: one for text-to-video and first or last frame to video generation, and one that generates video from a mix of reference images, video and audio. Output runs 4 to 15 seconds at 24 frames per second, up to 2K resolution, with 32kHz stereo audio and stable support for 11 languages, according to the repository.
Two supporting modules, a context-interpretation preprocessor and a 2K upscaling regenerator, are not open-sourced. The open weights are released under MiniMax's own H3 Community License rather than a standard permissive license, and support SGLang, vLLM, diffusers and ComfyUI for local deployment, the repository says.
For builders working with generative video, a 33 billion parameter model that natively ties audio to video generation, rather than bolting on a separate audio pass, is a smaller and more deployable option than the closed frontier video models it's aimed at competing with. The license terms are worth reading before building a product on top of it.