The video model that excludes the EU and the UK, but not Switzerland
The difference sounds like hair-splitting. It decides whether your media team may use a video model that costs nothing and runs in your own data centre.
In short
- HunyuanVideo is not banned in the EU. It is simply not licensed there. For the EU, the United Kingdom and South Korea, Tencent grants no right to use the model weights at all.
- In Switzerland the licence applies. The country is not on the list of exclusions, and there is no horizontal AI act here on the EU model.
- The trap is not the location but the group structure. An EU subsidiary, an EU data centre or a team in Berlin brings the exclusion back into the building.
Anyone who discussed generative video models in a Swiss media house this spring knows the line: 'we can't take Hunyuan, it's banned in Europe.' The line is well meant and wrong. No law bans HunyuanVideo. What is missing is something else. And that difference decides whether you may use the model.
Tencent publishes its Hunyuan models with open weights. Anyone can download them, run them locally, fine-tune them. What permits the use is not the law but a licence. And that licence draws a national border.
What the licence text actually says
In section 1(l), the Tencent Hunyuan Community License defines the term 'Territory', meaning the area for which any rights are granted in the first place:
Territory: worldwide
In the excluded territories Tencent grants no right to use the model weights
'Territory' shall mean the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea.
Tencent Hunyuan Community License Agreement, section 1(l)
And there is more: the licence text puts a notice in capitals above the same statement: THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA. That is not a footnote and not an oversight. Tencent has deliberately carved out three jurisdictions.
In passing, because it is easily missed on a first read: section 4 draws a commercial line as well. Anyone whose products together reach more than 100 million monthly active users needs a separate licence from Tencent. For Swiss media houses that is not a practical hurdle. For a platform business with a growth plan it is.
Why 'not licensed' is a different thing from 'banned'
A ban comes from an authority. It is set out in a law, it applies to everyone, and the sanction is a fine or a prohibition order. That is how the EU AI Act works: it bans certain practices and imposes duties on providers.
A licence exclusion is something else entirely. It is civil law. The model weights are protected subject matter that the rights holder controls. Copy them, adapt them or build them into a product without permission and you have no right of use. What follows then is not a supervisory measure but an injunction, damages, and in the worst case a product recall.
Ban (regulatory law)
An authority prohibits a practice by law. You learn of it from the official gazette, the rule applies to the whole market, and enforcement rests with the state.
not the case with HunyuanLicence exclusion (civil law)
The rights holder grants no rights for a territory. You learn of it only from the licence text, it concerns you alone as the licensee, and it is enforced by Tencent whenever Tencent chooses.
this is exactly the caseFor procurement the second case is the more awkward one. A ban is public, dated and the same for everyone. A licence exclusion sits in a text file in a repository, changes with the next model version, and nobody sends you a notification.
Why Switzerland is not on the list
The answer is unspectacular: the list in the licence text is exhaustive. It names the European Union, the United Kingdom and South Korea. Switzerland is not an EU member, so it falls under 'worldwide territory'. The rights apply.
That fits the regulatory picture. Switzerland has no horizontal AI act. On 12 February 2025 the Federal Council decided to ratify the Council of Europe's AI Convention and otherwise to regulate sector by sector: existing legislation is amended where AI creates concrete problems, rather than creating a framework act of its own. A consultation draft is announced for the end of 2026. Until then, and probably for a while beyond, what governs the use of media models in Switzerland is above all this: data protection law, copyright and personality rights.
Why Tencent carves out those three territories is stated nowhere. Our reading of it, expressly flagged as such: it is precautionary risk avoidance. The duties for general-purpose AI models under the EU AI Act (documentation, a copyright policy, a public summary of the training data) are hard to meet for a provider that simply puts weights on the internet, and harder still to document. Take the jurisdiction out of the licensed territory and the problem goes away. We cannot prove this; Tencent does not comment on it.
What the AI model in question looks like
Technically HunyuanVideo is not a curiosity but a serious model: text to video at usable quality, with open weights, and it runs on your own hardware. For media houses that do not want to go through an API, whether for data protection or for cost, that makes it one of the few serious options. Which is why the licence point weighs so heavily: it hits not a weak model but one of the few that really works on-prem, on your own hardware.
Stills from the official examples · Source: Tencent, aivideo.hunyuan.tencent.com
Under the surface it is a diffusion transformer with 13 billion parameters. Tencent combines a dual-stream block, in which text and image are processed separately at first, with a single-stream block that brings the two streams together. That is why the model holds image content and prompt together across longer sequences.
The trap is not where the company is registered
The obvious conclusion runs: we are a Swiss company, the licence applies to us, case closed. In most of the procurement conversations we have had on this, that is exactly the point where it got messy. Three situations bring the exclusion back into the building:
1. The subsidiary in the EU
If the model runs on the infrastructure of a German or Austrian group company, the use takes place there, that is, in the excluded territory. The parent company's registered office in Zug does not help.
2. The data centre across the border
'On-prem' in practice often means a rented GPU cluster. If it sits in Frankfurt, the same objection applies. With a model you want to run yourself precisely because of data sovereignty, the question of location arises anyway. Here it gains a second reason.
3. The distributed team
If the colleagues who actually operate the model sit in Berlin or London, they are using it there. With remote teams this is the case most easily missed.
What does stay unproblematic is the output. A video lawfully produced in Switzerland may be distributed in the EU. The licence governs the use of the model, not the distribution of the result. Distribution then raises the usual questions: labelling duties, personality rights, provenance. A different problem, the same care.
In the comparison matrix The row on the Hunyuan models carries this licence exclusion under 'DACH risk'. It also lists the on-prem alternatives that come without a territorial restriction. For video those are currently Wan 2.2 and LTX-2. To the matrix →What follows from this
The lesson reaches beyond Hunyuan: 'open weights' does not mean 'free to use'. Apache 2.0 and a community licence with a territorial exclusion are worlds apart, and the difference is not apparent when you download. It becomes apparent when somebody asks. By then the answer should be in the procurement file.
What we recommend
For video you run yourself, take Wan 2.2 or LTX-2. Both can be self-hosted and carry no territorial restriction, so that whole discussion falls away.
If it has to be Hunyuan because the quality decides it: settle where it will run before you procure, not afterwards. Data centre, cloud region and the location of the people operating it belong in the file in writing.
And as a principle: check the licence before the benchmark. A model you are not allowed to use does not need comparing.
For any AI model with open weights
- Read the licence text of the particular model version, not that of the model family. Providers change their terms between versions.
- Record in the procurement file which version you checked and when. A licence text in a repository is not a stable state.
- Keep an alternative ready that carries no territorial reservation. Then the discussion at the next licence change is a short one.
And if HunyuanVideo is specifically on the table
Then two further checks follow from the licence text:
Additionally for the Hunyuan models
- Check where the model runs, not where your company is registered. Data centre, cloud region, location of the people operating it.
- Check the 100-million-user threshold if there is a platform business behind it.
Open questions
Unclear whether you are allowed to use a model?
If you have a procurement ahead of you, or want to check whether a particular model is clean in licensing terms for your operation: in an initial conversation we go through your specific case, data centre, group structure and alternatives included. No sales pressure.
Sources
- Manufacturer. Tencent Hunyuan Community License Agreement, sections 1(l) and 4 · github.com/Tencent-Hunyuan/HunyuanVideo
- Switzerland. Federal Council, press release of 12 February 2025 on the review of AI regulation and the way forward · bakom.admin.ch
- Switzerland. Digital Switzerland, measure 'Draft for implementing the Council of Europe Framework Convention on Artificial Intelligence' · digital.swiss
- Manufacturer. Video, stills and architecture diagram from Tencent, official HunyuanVideo project page · aivideo.hunyuan.tencent.com. Reproduced to illustrate the model under discussion; rights with Tencent. The showreel was reduced to 1280 px for delivery and muted, otherwise unchanged.
- Manufacturer. Technical report on HunyuanVideo · hunyuanvideo.pdf
- Our own work. digitario, comparison matrix of AI media models · digitario.ch/ki-medienmodelle-vergleich