MiniMax releases omni-modal MiniMax-H3 on Hugging Face with native stereo audio and up to 2K video

MiniMax has published MiniMax-H3, a new model it describes as a general-purpose omni-modal generative system, on its official Hugging Face registry. The release lands as an image-text-to-video pipeline built on the diffusers library, meaning developers can load it directly through DiffusionPipeline.

Unlike a typical text-to-video model, H3 is designed to understand multimodal contexts composed of text, images, video, and audio, and to generate video with native stereo audio. According to the model card, output duration ranges from 4 to 15 seconds, the shorter side defaults to 768 pixels at 24 FPS, audio is produced at 32kHz stereo, and 2K generation is available through a dedicated H3-Regenerate-2K variant.

The specification covers a full video workflow: text-to-video, image-to-video, and video-to-video, as well as joint generation from text, image, video, or audio into synchronized audio-video output. For dialogue, the model provides stable support for 11 languages, including Chinese, English, Japanese, Korean, and Arabic.

The release ships several variants — H3-Base for general generation, H3-Context-IR for long-context understanding, and H3-Regenerate-2K for high-resolution regeneration — under the minimax-h3-community-license-agreement, an open-weight community license.

The registry entry appears to be fresh: the model card shows zero downloads and 148 likes, consistent with a same-day official update. Its tag set covers text-to-video, image-to-video, video-to-video, and text/image-to-audio-video generation.

What stands out is that native stereo audio is built into the generation pipeline, so picture and sound are produced together in one model rather than through separate post-production dubbing. For the open-source community, such all-in-one audio-video models remain rare, and the real generation quality plus the 2K regeneration workflow will be the key things to watch.

Next up: community reproductions and fine-tunes built on diffusers, how the H3 family performs on longer videos and multilingual prompts, and whether MiniMax opens up larger weights and an online API.