MiniMax H3
A 33 billion parameter open-weight video model that generates short clips with synced audio, but whose licence excludes the EU and other regions.
MiniMax H3 takes text, images, video and audio as one combined input and produces clips of four to fifteen seconds with native stereo sound, meaning the audio is generated together with the picture rather than stitched on afterwards. It ranked first for video editing with audio on Artificial Analysis when the weights went public in August 2026, and in the top three for text-to-video and image-to-video.
The licence is the part worth reading before the benchmarks. MiniMax’s Community Licence defines an applicable territory that explicitly excludes the European Union, the United Kingdom, the United States and South Korea from running the weights locally. The hosted API stays available everywhere. The terms also require prominent attribution, forbid training smaller models on H3’s output, and need separate written permission above 20 million dollars in yearly revenue.