MiniMax H3 and the small print: the best local video model Europe isn't allowed to use

On 3 August, MiniMax published the weights of its video model H3 (Hailuo 3.0) on Hugging Face — and technically this is a milestone: a 33-billion-parameter model that generates clips of 4 to 15 seconds at up to 2K/24fps, with native stereo audio, locally on your own hardware (Hugging Face).

In the Artificial Analysis rankings, H3 sits at number 1 for video editing with audio and in the top 3 for text-to-video. ComfyUI support arrived on release day.

In short: the thing many people have been waiting two years for — a “local Sora”. Realistically it needs a good 40 GB of VRAM, so workstation class, not laptop.

Then you read the licence

The “MiniMax H3 Community License” explicitly excludes the USA, the EU, the United Kingdom and South Korea from the permitted territory (TechTimes). If you are based there, you may neither run nor modify the weights, nor use their outputs.

Which means we may not show any MiniMax videos we produced ourselves in this article either.

Important for context: the block is legal, not technical. Anyone can download the model and try it out; it simply is not licensed here. The hosted MiniMax API remains available worldwide, and MiniMax grants individual licences on request to users who can demonstrate content compliance.

For a weekend of tinkering that is a theoretical problem. For enterprise use it is a real one. What US export control was for Fable in June, a Chinese licence clause is here. Access to intelligence is being steered geopolitically.

Is there a clean way to use Chinese open-weights models anyway?

Mistral sketched an answer in August. On 11 August, the French company announced it would offer third-party models such as Z.ai’s GLM-5.2 on its own European infrastructure, with a selectable inference region (EU or US), an SLA-backed priority tier and regional data control (Mistral).

On top of that comes a coalition with Amadeus, ASML, Capgemini, CMA CGM and the Caisse des Dépôts, which is meant to secure European compute capacity long term via “European Compute Units” — reportedly up to 1 GW by 2030. The motto, put bluntly: Chinese models, hosted in Europe, secured by contract.

Our take

“Open weights” is not a legal term. Read the licence before you plan. There are worlds between Apache 2.0 (Inkling, for instance), community licences with user thresholds (Llama style) and territorial exclusions (H3). For enterprise use, licence review now belongs in every model evaluation, on an equal footing with benchmarks.

The capability is there anyway — and it diffuses. Even though H3 is blocked for Europe, the “local Sora” level has now been reached, and it will be matched under usable licences. If you have video use cases, build the architecture today so that the video model is replaceable. Our GenAI media conclusion from June still stands: orchestration is the lever, the model is a building block.

Sovereignty is becoming a product. Mistral’s package — choose your region, SLA, open third-party models — is exactly the kind of offer we have considered overdue for months. Whether the ECU coalition delivers the promised capacity remains to be seen. But as a contractual model, “sovereign consumption of open models” is now a real third option between US cloud and self-hosting.

Planning video use cases and unsure which models you may actually use? Let’s talk.