The open-weight licence trap

MiniMax H3's open-weight licence excludes the United States, the European Union, the United Kingdom and South Korea. That single clause removes the best-value model on the board from most Western brand work, and it has nothing to do with how good the model is. Practitioner notes, not legal advice.

H3, also called Hailuo 3.0, was announced on 31 July 2026. It ranks top-three for text-to-video on Artificial Analysis and leads the board on on-screen text rendering, which matters more than it sounds: legible type in frame is the thing that usually forces a model-generated shot back into After Effects. H3 Max leads spatio-temporal consistency. The specs are generous for the price: up to 2K at 24fps, clips of four to fifteen seconds, native stereo audio, and reference-to-video that accepts nine images, three video clips and three audio clips at once. The model overview.

And it costs $4.80 per minute at 768p, roughly a seventh of Seedance 2.5. For anyone making volume work rather than a single hero film, that is the number that decides the quarter.

The trap

There are two MiniMax H3s as far as your contract is concerned.

The open weights carry a licence that excludes the US, the EU, the UK and South Korea, and requires separate authorisation above $20M in annual revenue. If you are a studio in Singapore making a film for a German client that will run in Europe, the territorial clause is not a technicality you can work around with a VPN and a shrug. It is the clause your client's procurement team will find.

The paid API has no territorial exclusion.

So the route decides the answer, and the route is invisible in the finished file. Two shots can look identical, come out of the same model, and sit on opposite sides of a licence boundary depending on whether someone ran local weights or called the API. That is why the clearance record has to name the route and not just the model.

Pika H3 Flex

Pika shipped H3 Flex on 29 October 2026: a post-trained MiniMax H3, sold as a commercial product, with open weights. It inherits the same question, and how Pika resolves it could not be verified. We have not found terms that state whether the territorial exclusion carries through to Flex, is replaced, or is licensed around. Until that is answered in writing by Pika, treat Flex as unresolved for US, EU, UK and South Korean work rather than assuming either outcome. Ask them directly and keep the reply.

The other open-weight terms worth knowing

• LTX 2.3 open weights. Free under $10M revenue. AI-content disclosure is required, which means the licence itself obliges you to label, independent of any platform rule or statute.
• Alibaba Wan 2.2 open weights. Apache-2.0, which is genuinely clean for commercial use and the easiest approval you will get all year. One catch sits next to it: outputs from the Alibaba Model Studio trial are generally restricted from commercial use. Same family, different artefact, different answer.
• Wan 3.0. API only, closed weights, no indemnification. The newer model is the less flexible licence, which is the opposite of what most people assume.

Why this stopped being academic

In 2025 the open models were a clear tier behind the closed ones, and the licence question was a hobbyist concern. In 2026 that gap closed. A fal post-train of MiniMax H3 topped the silent Artificial Analysis board. Pika shipped a commercial product built on open weights. The frontier now includes models whose terms were written for research distribution rather than for a brand film running in forty markets.

“The cheapest capable model in the world is worth nothing to you if its licence excludes the market you are selling into.”

Which produces a specific failure that we have watched happen to other people. A team picks a model on quality and price, builds a pipeline around it over several weeks, trains everyone on it, and then discovers at contract stage that the licence excludes the market the campaign is selling into. The model works. The pipeline works. The work cannot ship. Nothing in the benchmark would have warned them, because benchmarks measure output and licences govern use, and no board on the internet reconciles the two for you.

Check the territory clause before the benchmark

Four questions, answerable in about twenty minutes per model, before any pipeline work starts:

• Which route? Open weights and paid API are different licences on the same model. Decide which one you are building on and write it down.
• Which territories are excluded? Check against the markets of release, not the market you are producing in. A Singapore studio is not exempt from a clause that governs where the work runs.
• Is there a revenue threshold? MiniMax at $20M, LTX at $10M. Whose revenue, the studio's or the client's, is worth asking explicitly rather than assuming the convenient reading.
• Does the licence itself demand disclosure? LTX does. A licence-level disclosure duty applies even where the platform and the statute would not require one.

The cheapest capable model in the world is worth nothing to you if its licence excludes the market you are selling into. Put the territory clause at the top of the model evaluation sheet, above the quality score, and the problem disappears before it costs anything.