MiniMax has officially open-sourced its H3 video model. You can now run the weights locally to generate 768p video — but full 2K output still requires the official API.
The release includes two H3-Base variants: one for text-to-video and first/last-frame generation, and another for complex generation and editing using reference people, actions, videos, and audio. You only need to load the one that matches your task, not both.
H3-Base's generation backbone is a 33-billion-parameter dense single-stream Transformer, with an encoder based on Qwen3-VL-32B. Running locally, it produces 4 to 15-second 768p videos with stereo sound.
This release isn't the full H3 system. The part responsible for understanding complex material relationships...