MiniMax opens the weights of a 2K video model that generates its own stereo soundtrack

Three days after launching H3 on its consumer Hailuo platform, MiniMax published the model's weights on Hugging Face and ModelScope under its own Community Licence. The company frames the move as a correction of an imbalance: language models iterate in the open, while video generation has been dominated by closed systems with slower cycles and no ecosystem around them. What is released is a 33-billion-parameter dense transformer, H3-Base, paired with a frozen Qwen3-VL-32B encoder, plus the module that rebuilds its 768-pixel draft at 2K. Output runs 4 to 15 seconds at 24 fps with 32 kHz stereo audio, in aspect ratios from 21:9 to 9:16, with spoken dialogue stable in eleven languages. The model reads text, images, video and audio as one context: a prompt can ask it to borrow a camera move from one clip, a face from an image and a voice from an audio file, described in ordinary language rather than selected from a fixed task menu. One part stayed closed. H3-Context-IR, the layer that interprets those mixed inputs and compiles them into the structured form the base model expects, depends on hosted models and was not open-sourced. MiniMax calls it critical to output quality and offers an API for it, along with a prompting guide for anyone building their own substitute — so a fully self-hosted pipeline is possible, but not with the vendor's own orchestration. On MiniMax's platform the hosted route costs $0.13 per second of 2K video and $0.08 at 768p, with generated audio billed at nothing.