Three Billion Euros for Sovereign Open Weights — What Does That Buy a Small Business?

What happened

The round is 3 billion euros, and the valuation after it is above 21 billion euros. Samsung Electronics led the investment, co-led by the EQT-managed Scaleup Europe Fund and PSG Equity. Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg came in as new investors; the existing ones include ASML, Bpifrance, NVIDIA and Salesforce Ventures.

In the company's own wording this is "the largest equity fundraising round ever completed by a European technology company", three years after it was founded. The announcement names 20 countries and more than 125 enterprise customers, among them Airbus, ASML and HSBC.

A single sentence carries the strategy. The whole chain — "open-weight models, the infrastructure and the compute capacity they run on, and the products that bring them into production" — is kept in one hand so that "customers are never locked into a single vendor's roadmap, pricing or availability."

Worth noting too is what the announcement does not contain: no EuroHPC, no EU co-funding, no AI gigafactory, no IPCEI. This is a private company on private money.

What "open weights" means here

Not one single thing. With downloadable models the licence decides what you are allowed to do — and the licences sit on four rungs:

Apache 2.0    free, commercial too
Modified MIT  free, up to a revenue cap
CC BY-NC      commercial use NOT allowed
Llama licence the publisher's own terms

In Mistral's own catalogue it looks like this. Under Apache 2.0 you can download Mistral Large 3 (675 billion parameters, 41 billion of them active), Mistral Small 4, the three sizes of Ministral 3, Devstral Small 2 and part of the Voxtral audio family. Under a modified MIT licence sit Mistral Medium 3.5 and the large Devstral: you may use those freely as long as your company's global consolidated revenue in the previous month did not cross 20 million dollars — above that you need a commercial licence. The Voxtral text-to-speech model, on the other hand, is CC BY-NC, so it cannot be used commercially. And some models are reachable only through the API: OCR, Codestral and the embedding models.

So the "open-weight company" label on its own does not answer the question of what you may do with a given model. The licence line on the model card answers it.

What "sovereign" means — the fine print

Here too it pays to read down to the documentation. Mistral made its regional endpoints generally available in August: there is a global one, an EU one and a US one. What the documentation states:

• the EU endpoint processes the request in EU and EFTA data centres;
• the regional endpoint costs 1.1 times the list price — so the price of sovereignty here is a concrete 10% surcharge;
• the control plane is not regional: account settings, API keys, billing and usage statistics can leave the region you picked;
• the agents, batch and files APIs do not work on the regional endpoint;
• zero data retention takes a separate request, only on a paid plan, and it covers the stateless endpoints only — not the web interface, not the agents, not file handling.

On the infrastructure side the commitment is 1 GW of European capacity by 2030; the Swedish data centre in Borlänge is a 1.2 billion euro investment opening in 2027, and the French site takes 18,000 NVIDIA systems in its first phase. These are real, documented plans — but in September 2026 they are still plans.

What open weights actually buy

The most useful part is neither the licence nor the sovereignty, but something far more mundane: weights you have downloaded cannot be retired out from under you.

Mistral's own model lifecycle policy gives 6 months of notice on generally available models and 1 month on experimental and preview ones, and it states the end plainly: "Once retired, requests to its identifiers fail with a 404 error." This is not theory: several of their models were retired during 2026 — Mistral Small 3.2 and an earlier release of Devstral 2 stopped answering in the API on 31 July.

The weights of those same models are still downloadable today. That is the difference in practice: the API version can go away, the file does not. If a process of yours needs stable output for years — because another system is built on the format of that output, say — then this difference is worth more than a few percent of quality.

In exchange, open weights do not run for free. The 8B version of Ministral 3 in FP8 fits into 12 GB of video memory, so it starts up on a single better consumer card. Mistral Large 3 in FP8, though, asks for a whole server node. "I can download it" and "I can actually run it" are two separate questions.

Which ones handle the smaller languages

Here comes the cold shower: Mistral's model cards do not list Hungarian. Large 3 and Ministral 3 enumerate eleven languages, Small 4 and Medium 3.5 twenty-five — Spanish is in all of them, Hungarian in none. The prose text mentions "dozens of languages", without a list.

There are, however, several European open-weight models that do spell Hungarian out:

One caveat on ALIA: Hungarian is in the language list, but the post-training concentrated on Spanish, Catalan, Basque, Galician and English, and the model card itself says that "performance in those additional languages is not guaranteed". So "supported language" here too means try it on your own texts. More on that in which models handle smaller languages — the measuring method is in there as well.

What to do now

1. Read the licence line, not the headline. Before you build on an open-weight model, check on the model card what the licence is, and whether there is a revenue cap or a non-commercial clause.
2. If the data cannot leave the EU, read the fine print. A regional endpoint is not the same thing as the whole service running regionally — the control plane, file handling and the agents are separate questions. That is what European AI options covers, and its checklist is worth going through.
3. Download one small model, do not just plan to. An 8 billion parameter model fits on a stronger laptop or a cheap server. Finding out how big the quality gap really is on your own task is worth more than any comparison table.
4. Keep the option, not the loyalty. The biggest practical benefit of open-weight models is not the price but that your plan B survives. For that it is enough if your system can switch models through a single configuration line.

Common misconceptions

• "Open weights are free." Downloading the weights is free. Running them asks for hardware, operations and electricity — and on the large models that is a serious line item.
• "Open weights mean I can do anything with them." Only under Apache 2.0 or a similarly permissive licence. There are non-commercial licences, and licences tied to a revenue cap.
• "A European model is automatically better in a smaller language." It is not. Mistral's model cards do not even list Hungarian, while other European, research-funded models do.
• "If the provider is European, the data stays in Europe too." Only if you use the regional endpoint, and you know what it does not include.
• "Sovereignty comes for free." The regional endpoint here runs with a 10% surcharge, and a few features are not available on it.

What to take away

A three billion euro round is, on its own, investor news. What translates from it into everyday work is three concrete things: the licence says what you are allowed to do, downloaded weights cannot be retired out from under you, and the price and the limits of the "sovereign" endpoint are written out in the documentation. Those three together are worth far more than the word sovereignty. And if a smaller language matters in your work, the largest European company is not necessarily the best answer — whereas on the list of models built from EU research funding, Hungarian is there.