According to CoinMeta, the MiniMax team announced that the H3 video model has been officially open-sourced, allowing users to generate 768p videos locally. However, generating complete 2K videos still relies on the official API. This release includes two sets of H3-base models: one set is used for text-to-video generation and the creation of opening and closing frames, while the other set is used for complex generation and editing involving reference characters, actions, videos, and sounds. The core of H3-base's generation is a dense single-stream transformer with 33 billion parameters, and the encoder is based on qwen3-vl-32b. The team revealed that they plan to open up the 2K video module in the future and will introduce acceleration solutions to improve local running efficiency. Currently, H3 only supports 768p video generation locally, and the team acknowledges that there are issues such as blurry images when generating videos by referring to pictures, videos, and sounds simultaneously; they are working on optimizations to address these problems.