MiniMax H3 Officially Open Source: Local 768p Video Generation Available, Full 2K Still Relies on Official API

BlockBeats News, August 3, MiniMax officially released the weight of the video model H3. This release includes two sets of H3-Base: one set is responsible for text-based video and keyframe-based video, and the other set is responsible for referencing images, videos, and audio simultaneously. Users can choose one set based on the task without the need to load both simultaneously.

H3-Base is a dense model with 330 billion parameters. Local deployment can generate 4 to 15-second, 768p videos with stereo sound, and supports SGLang, vLLM, Diffusers, and ComfyUI.

This release does not include the complete H3 system. The H3-Context-IR responsible for understanding complex material relationships, the H3-Regenerate-2K to regenerate results to 2K resolution, and sparse attention implementations have not been released. Developers still need to call the MiniMax API to reproduce the official full 2K effect.

There are also significant restrictions on licensing. The MiniMax H3 Community License is not applicable to the United States, the European Union, the United Kingdom, and South Korea; authorization needs to be applied for separately in these regions. For commercial products with annual revenue exceeding $20 million, prior written permission from MiniMax is also required.

Click the original text link below to join the BlockBeats · Feishu AI News Channel, monitoring global AI hot topics and news 24/7.