MiniMax-H3

MiniMax-H3 is a generative model developed by the Chinese company MiniMax that produces video with natively synchronized stereo audio. It accepts omni-modal input — text, images, video clips and audio — and generates short video clips with a matching audio track.

The system core (H3-Base) is a 33B-parameter dense, single-stream Transformer (with roughly 13B parameters in AdaLN-related branches) that uses separate visual and audio latents (VisualVAE and AudioVAE). The architecture comprises three modules: H3-Context-IR (preprocessing and orchestration), H3-Base (768p generation) and H3-Regenerate-2K (regeneration up to 2K).

The model generates video of 4–15 seconds at 24 FPS with 32 kHz stereo audio, at resolutions up to 2K and aspect ratios 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Weights are available on Hugging Face under the MiniMax H3 Community License Agreement; the model is also accessible via API (platform.minimax.io) and the Hailuo app (hailuoai.video), and can run locally through diffusers, vLLM, SGLang and ComfyUI.