Overview
- MiniMax published the H3 model and documentation on Hugging Face under the MiniMax H3 Community License, making the code and weights available to the public.
- H3 is a task‑general multimodal generator that accepts text, images, video and audio as context and produces short videos of 4 to 15 seconds with native stereo audio.
- The system is split into three cooperating modules: H3-Context-IR for producing a context representation, H3-Base for creating 768p outputs, and H3-Regenerate-2K for re‑generating higher‑detail 2K results.
- Two input modes manage references and limits: H3-Base-FL2VA accepts 0–2 images for text→video or first/last frame guidance, while H3-Base-Ref2VA accepts up to 9 images, up to 3 video segments (each 2–15s, total ≤15s) and up to 3 audio segments that must accompany image/video inputs.
- Official output specs include 24 fps video, 32 kHz stereo audio, multiple aspect ratios and stable dialogue support for 11 languages, which should let creators and developers build multilingual short‑form video tools and experiments.