Shanghai lab MiniMax published the weights for H3, its 33-billion-parameter video model, to Hugging Face on August 3, then bolted on a community license that carves out the United States, the European Union, the United Kingdom and South Korea. Four of the world’s largest developer markets can look at the model card. They can’t legally run the weights.
The license, effective August 2, forbids local deployment, modification, or redistribution of H3 by users in those jurisdictions under Sections I.3, I.5 and V.4. Developers there “must apply for a formal license,” per the South China Morning Post, which MiniMax says it’ll grant after reviewing deployment scenarios and compliance controls. Open weights, gated by geography and paperwork.
The model itself is aggressive. H3 produces up to 15 seconds of video at 2K resolution with native stereo audio, and handles cross-modal prompts spanning text, images, video and reference audio in a single pass. Artificial Analysis ranks it first globally in video editing, second in text-to-video, and third in image-to-video. The API lists at CNY¥0.8 per second, or roughly USD$1.95 for a 15-second 2K clip, which OpenSourceForU pegs at about one-twelfth the price of ByteDance’s Seedance 2.5. Investors noticed. MiniMax Group (0100.HK) closed up more than 10% at HKD$249.40 on the Hong Kong exchange.
MiniMax’s own explanation, posted in a Hugging Face Q&A alongside the release, cites two forces. Certain jurisdictions “are currently developing or enforcing AI-related regulations that may have specific implications for generative video models,” and the company is “involved in ongoing copyright-related legal proceedings specifically concerning generative video AI.” Asked whether US developers would ever get access, MiniMax answered: “not yet, not not ever.”
That litigation isn’t abstract. On September 16, 2025, Disney Enterprises, Universal City Studios and Warner Bros. Discovery sued MiniMax, its US-facing operator Nanonoble Pte. Ltd., and its parent in the Central District of California, alleging that MiniMax’s Hailuo platform was trained on unauthorized copies of characters including Spider-Man, Darth Vader and Wonder Woman. US District Judge Stanley Blumenfeld denied MiniMax’s motion to dismiss on May 26. Ryan Lee, MiniMax’s Head of Developer Relations, confirmed on KuCoin that the US carve-out follows directly from the Hollywood case.
The strategic logic is legible. Ship the weights, capture the mindshare of every developer outside a courtroom’s reach, and turn the excluded markets into a bilateral licensing pipeline the company can staff with lawyers rather than engineers. It’s an open-source posture routed around the parts of the world where open-source infringement claims actually get litigated.
The backdrop tightens the frame. Andersen v. Stability AI goes to trial September 8 in the Northern District of California, the first federal test of whether AI-generated visual output constitutes an infringing derivative work. MiniMax has decided not to wait for the answer.
Sources
- https://www.minimax.io/blog/minimax-h3
- https://huggingface.co/MiniMaxAI/MiniMax-H3/discussions/12
- https://www.scmp.com/tech/tech-trends/article/3362951/chinas-minimax-curbs-overseas-access-new-ai-video-model-over-copyright-disputes
- https://www.techtimes.com/articles/322904/20260804/minimax-h3-open-weights-exclude-us-eu-uk-korea-local-deployment.htm
- https://www.opensourceforu.com/2026/08/minimax-releases-h3-multimodal-ai/