MiniMax releases 33B-parameter H3 video model weights, but keeps context module and 2K upscaler closed
AI video model releases
MiniMax releases 33B-parameter H3 video model weights, but keeps context module and 2K upscaler closed
MiniMax released open weights for its H3 video model on August 3, 2026, carrying 33 billion parameters, two days after AI Insiders reported the company had announced open weights without a download link or parameter count.
On Artificial Analysis rankings, H3 placed first in Video Editing, second in Text-to-Video, and third in Image-to-Video, making it the first open-weight model to top a category on that leaderboard.
H3 generates 4 to 15 second clips with stereo sound from combined image, audio, video, and text inputs, accepting up to nine reference images plus three video clips and three audio clips per prompt.
Two components stay closed: H3-Context-IR, which formats prompts and reference media before rendering, and the 2K resolution module; running H3 locally via ComfyUI caps output at 768p versus the higher resolution available through MiniMax's hosted service.
The open license limits commercial use to companies with $20 million or less in annual revenue, and the same day ByteDance shipped its closed Seedance 2.5 model producing 30-second clips with built-in audio, double H3's max length.
cho.sh