MiniMax opens the weights of H3, its video-and-sound model

MiniMax launched H3 on 31 July and opened its base weights on 3 August, on Hugging Face and ModelScope.

Two checkpoints

The open release has two, both in BF16. FL2VA generates from text and up to two images — a first frame, a last frame, or both. Ref2VA takes references: up to nine images, three videos of 2 to 15 seconds and three audio files, twelve files in all. Two further variants, H3-Context-IR and H3-Regenerate-2K, stay on MiniMax's own service.

What it makes

Clips of 4 to 15 seconds at 24 fps with 32 kHz stereo sound, in six aspect ratios from 21:9 to 9:16. Characters can speak in eleven languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. The open base generates at 768p.

Under the hood

A 33-billion-parameter dense single-stream transformer, with Qwen3-VL-32B reading the prompt. It runs on SGLang, vLLM, diffusers and ComfyUI; MiniMax's SGLang example uses four GPUs, and MiniMax says sparse attention is coming later.

Who may use the weights

The MiniMax H3 Community License applies worldwide except the European Union, the United Kingdom, South Korea and the United States. MiniMax says that is about regulation still being settled there, invites organisations in those places to apply for a licence, and keeps its API open globally. Where the licence does apply, commercial products with more than $20 million in yearly revenue need MiniMax's written authorisation, and any commercial product built on H3 must show "MiniMax H3" in its interface.