MiniMax open-sources new video model H3 with 2K resolution and native stereo audio

MiniMax has officially open-sourced its new general video model, H3. This model can understand text, images, video, and audio together, and generate videos up to 2K resolution, 15 seconds long, with native stereo sound. It was trained to handle complex multimodal instructions from the start.

H3 offers two input modes: one for text-to-video or first/last frame generation, and another that accepts up to 9 images, 3 video clips, and 3 audio clips, with a total of 12 inputs. Output supports various aspect ratios, default 768p resolution, 24 FPS, and 32 kHz stereo audio, covering 11 languages.

The full H3 system includes three parts: H3-Context-IR for understanding instructions, H3-Base for generating 768p video, and H3-Regenerate-2K for upscaling to 2K with better details. The model is released under a community license, and the source code is available online.