Why Mistral and Qwen are champions of European sovereignty | europeanGPU
Understand it, without jargon

Why Mistral and Qwen are champions of European sovereignty

Ministral (built in France by Mistral AI) and Qwen are open-weights models. This technical characteristic has a very concrete consequence: your data never flows back to the model’s publisher. This page explains why, step by step, with no technical background required.

Step 1 — A model is a file

A large language model (LLM) is not a “connected robot”. It is the result of a long training process, condensed into one very large file of numbers: the weights. Think of an encyclopaedia carved in stone: once written, it never changes, it speaks to no one, and it has no internet connection. To use it, you need a powerful computer (GPUs) that reads the file and computes the answers.

The whole question of sovereignty then comes down to one sentence: where is the computer that reads the file, and who owns it?

Step 2 — Two very different worlds
Closed proprietary model
Your application
contracts, patient records, emails…
↓  your questions AND your documents leave  ↓
The model publisher’s servers
often outside Europe, under foreign law
↓  the answer comes back… and your data stays  ↓
Logs, copies, possible reuse for training: the publisher decides.
Open-weights, at europeanGPU
The publisher releases the weights (Ministral, Qwen…)
one file, downloaded once
↓  the file travels once, then nothing more  ↓
europeanGPU GPUs, in Europe
the model runs here, cut off from its publisher
↕  your questions and answers travel only here  ↕
Your application
nothing ever flows back to the publisher

The decisive point: in the open-weights world, the publisher supplies the file before any use, then disappears from the loop. It sees neither your questions, nor your documents, nor the answers — it does not even know you are using the model.

Step 3 — The questions we get asked

Can the model “phone home” to its publisher?

No. Weights are numbers, not a connected program: they have no network access, no automatic updates, no “black box” that communicates. Our inference servers are isolated in any case: nothing can leave for the outside.

Can the publisher cut off our access?

No, again. The file is already with us, under a permanent licence. A commercial or geopolitical decision by the publisher cannot switch off the service — the exact opposite of dependence on a proprietary API.

Are we allowed to do this?

Yes: Ministral and Qwen are released under permissive open licences, which explicitly permit commercial use, hosting and modification. That is a deliberate choice by the publisher.

Can we verify what is running?

Yes. Since the weights are public, your technical team can download the same file, compare the fingerprints and reproduce our results. Trust does not rest on our word, but on a verifiable file.

And Mistral, on top of that?

Ministral adds a further layer of sovereignty: the publisher itself is European. Open weights + a French publisher + European hosting = a sovereign chain from end to end, with no extraterritorial link anywhere.

“An open-weights model operated in Europe is a permanent import of know-how — not a revocable subscription to a foreign intelligence.”
— Why we chose Mistral and Qwen

Want to see the data sheet?

Capabilities, benchmarks and integration examples are on the Qwen3.8-27B and Ministral 3 14B pages. And for your security teams: why Mistral and Qwen are an answer to cybersecurity challenges →

Book a meeting