Mistral’s largest model runs on European servers under European law. The weights follow this month.

On 6 October Mistral AI opened a public preview of Mistral Large 4, the largest model the French company has built. Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centres in Europe, and that the preview is served from the same hardware. The API is open now through Mistral Studio. The model weights are due by the end of October.

The headline figures are 1 trillion parameters in total and 49 billion active on any single request. In plain terms, only a slice of the model works on each question, which keeps the cost per answer down. The model reads images, charts and documents as well as text. Mistral also says a large share of its training data was multilingual and spanned more than 160 languages, including every official language of the European Union. Estonian is on that list. Preview pricing on Mistral’s site is 1.36 US dollars per million input tokens and 4.18 per million output tokens.

Why European matters more than largest

One line in the announcement deserves more attention than any benchmark. Mistral says it will offer a European deployment that it operates end to end, independently of other digital service providers and under European law. Until now, a company that wanted a top-tier model usually sent its data to an American provider, or to a European cloud hosting an American model on the American company’s terms. A strong model trained, hosted and governed in Europe removes a reason many Estonian firms have given for keeping AI away from their most useful data.

Once the weights are public, the model can also run on a private cloud or on your own servers. Mistral makes that case for security teams, where losing access to a model in the middle of an incident is itself a risk. The same logic applies to a law firm with client files, a clinic with patient records, or a manufacturer with drawings it will not upload anywhere. Last week we wrote about Metrosert testing AI on health data that never leaves the country. The principle is the same, and this model brings it within reach of companies without a research budget.

What Mistral claims, and what it has not shown yet

By Mistral’s own numbers the model sits among the strongest open-weight models in coding, cybersecurity, finance and legal work, and in some image tasks ahead of closed models. Those are the vendor’s figures. Independent results will arrive once the weights are public and other teams can run their own tests. The announcement does not state the licence terms for the weights, and Mistral says the training run is still in progress and that architecture details and more benchmarks will follow with the release. Treat the preview as a preview.

One more point from the announcement. Until the weights ship, Mistral is red-teaming the model with cybersecurity partners and state authorities, and the public version carries the usual safeguards. The company notes that the model refuses malicious cybersecurity requests more often than other open models it compared against. For a buyer this is useful: it says something about how the model behaves when an employee pastes in a request they should not.

What an Estonian company should do this week

• ✳Take out the list of tasks you ruled out because the data could not leave your control. Each one now has a candidate model to test against
• ✳Have a developer run your own test set through the preview API: your documents, in Estonian, with the questions your staff really ask. Score the answers yourself and ignore the benchmark charts
• ✳Put the preview price on the same spreadsheet as what you pay today per million tokens, and note which deployment region your current provider actually uses
• ✳Ask your IT team or cloud partner whether they could host a one-trillion-parameter open-weight model when it ships. For most companies the honest answer is no, and the European-hosted API is the realistic route
• ✳Mark the end of October in the calendar. By then the weights and their licence should be public, and that is the point to decide, not now

The practical consequence is simple. The sentence “no, that data cannot leave” no longer ends the conversation about AI in your company. A capable model exists that was trained in Europe, is served in Europe and will soon run on hardware you control. The question for Monday is which task you test first.