How do you use Mistral models when your data must stay in Europe?

"Our data must stay in Europe" is a common requirement for clinics, law firms, public bodies and many companies that handle customer records. It usually means two things: personal data is processed under European rules such as the GDPR (the EU's General Data Protection Regulation), and it is stored and processed on servers in a place you can name. Mistral AI, based in Paris, is often on the shortlist because it offers several ways to use the same models, from a chat app to software you run yourself.

The useful question is not "Is Mistral European?" but "Which route keeps this data where it must be?" Each route below gives a different answer.

Four ways to use Mistral, from least to most control

1. Le Chat, the chat app. Mistral's terms say Le Chat keeps your prompts and replies as chat history and, on the free service, may use them to improve its models. Paid subscribers can opt out of training, but the opt-out only covers what you send afterwards. The terms also ask users not to enter sensitive personal data or confidential business information, and warn that anyone with a shared conversation link can read it. Le Chat suits drafts and general questions, not client files.

2. The API on La Plateforme. An API (application programming interface) lets your own software send text to a model and get an answer back. Mistral's terms are governed by the GDPR and French data protection law, and for the API, Mistral says it does not train its models on customer data. The terms do say prompts and outputs are kept for 30 days to check for misuse, and that customers with a legitimate reason can ask for zero data retention, which Mistral may grant or refuse. Read that section and sign the data processing agreement before going further.

3. Through a cloud you already use. Mistral Medium 3, released on May 7, is available on Amazon SageMaker as well as La Plateforme, and Mistral has announced it for Google Cloud Vertex, Azure AI Foundry, IBM watsonx and NVIDIA NIM. Mistral Small 3.1 is already on Vertex AI. If your organization has a European region and a contract with one of these clouds, this route can keep processing inside an agreement you already trust. Check that the model is actually offered in the region you need.

4. On servers you control. This is the strongest answer to "where is the data?" Mistral Small 3.1, released in March under the open Apache 2.0 license, can be downloaded and, according to Mistral, runs on a single RTX 4090 graphics card or a Mac with 32 GB of memory. Mistral Medium 3 can be self-hosted on four or more GPUs, on premises or in a private cloud, by arrangement with Mistral. Nothing leaves your building or your chosen data centre unless you send it.

A worked selection

Here is how three organizations might choose. The cases are illustrative.

Notice that the clinic picks the smaller model. That is a data decision, not a quality verdict: it can run Small 3.1 on its own hardware today, test it on real (properly protected) messages, and move up only if results fall short.

Questions to settle in writing

Whichever route you choose, get clear answers before personal data goes in:

• Where are prompts processed and stored, and for how long?
• Is any of it used for training, and how is that switched off?
• Which subcontractors handle the data, and in which countries?
• How do you delete data, and how do you prove it?

A vendor's marketing page is not enough here. The answers belong in the contract or the data processing agreement, and your own legal or privacy adviser should read them.

Which Mistral model fits your budget?

For occasional use, the API is the cheapest way in because you pay only for what you send. Mistral prices Medium 3 at $0.40 per million input tokens and $2 per million output tokens (a token is a piece of a word, roughly three-quarters of an English word). As an illustration, summarizing 1,000 letters a month at about 2,000 input tokens and 500 output tokens each comes to 2 million input tokens ($0.80) and half a million output tokens ($1), under $2 a month before any taxes or platform fees.

Self-hosting changes the bill from per-word to per-machine. Mistral Small 3.1 has no licence fee, so the cost is a capable workstation or server, the electricity and someone to maintain it. That makes sense when the data must stay in-house or the volume is steady and high. Medium 3 on your own infrastructure needs at least four GPUs and a conversation with Mistral, which puts it in enterprise territory.

A practical path: prototype with the API on non-sensitive sample data to learn what quality you need, then decide whether Small 3.1 on your own hardware is good enough for the real data, or whether Medium 3 in a private deployment is worth the larger investment.