MiniMax releases H3 weights

MiniMax has released the weights for H3, a 33B-scale omnimodal model that processes text, image, video, and audio within a single context and simultaneously generates video and stereo audio.

H3 generates video up to 2K resolution, 15 seconds long, at 24fps, producing dialogue, sound effects, and background music together within the same denoising loop without separate post-processing. Its architecture unifies audio, sound effects, and music, text/image/video inputs, and multi-shot, reference, and editing tasks into a single Omni Transformer.

Two types of 33B checkpoints have been released on Hugging Face, and the following ecosystems supported the release from day one.

• SGLang
• vLLM
• Hugging Face Diffusers
• ComfyUI

However, the MiniMax H3 Community License designates the Republic of Korea, along with the United States, the European Union, and the United Kingdom, as 'Excluded Territories.' As a result, downloading, distributing, or operating the weights within Korea is not covered by the default license permissions, and separate approval is required.

There is no technical download restriction on the repository, but usability should be judged based on the license terms, not on download availability. MiniMax is accepting approval applications from organizations in restricted territories that have compliance controls and safeguards in place, and within Korea, the currently unrestricted API remains the basic path for usage.