Mistral Large 4: the European model you can already try via API, and when it makes sense to

On 6 October 2026 Mistral opened the preview of Large 4, its largest model, trained from scratch on 3800 NVIDIA GPUs in its own datacentres in Europe ¹. If you have ever wondered whether there is a serious European alternative to ChatGPT, this is the closest thing to a yes seen so far.

The short answer to your question: for many office jobs, yes, it's good enough. For the hardest code, the best American models are still a step ahead. And for now you are using a preview, not a finished product.

What Mistral announced, without the jargon

Large 4 has 1 trillion parameters, of which 52 billion are active for each response ¹. In plain terms: it's a huge model, but built so that it doesn't use all that power every time.

It reads text and images together, including documents, charts and technical drawings. It was trained on more than 160 languages, including all the official languages of the European Union ¹. If you work in Italian, that is not a detail.

Today you use it via API on Mistral Studio. The weights, meaning the actual model that you can download and run wherever you like, Mistral promises to release by the end of the month ¹.

How good is it really, compared with the Americans

Honesty is needed here: all the figures that follow come from Mistral's announcement, not from a third party that has reproduced them from scratch.

On code, Large 4 solves 61.7% of DeepSWE problems ¹. In a blind evaluation run with Surge AI, where professional programmers voted without knowing which model they were looking at, it came second out of five, behind only Claude Opus 5 ¹. A good result, but first place is still American.

On business workflows it does better. On AutomationBench, which measures 657 processes across Gmail, Google Sheets, Slack and Salesforce, it scores 59.9% and beats several Chinese competitors ¹.

On images, Mistral claims its strongest result: on a visual recognition test it reports 42% against 41% for GPT-6 Astra ¹. One point of difference. Interesting, not decisive.

What "keeping data in Europe" means

This is the part that matters most for you, and also the most concrete.

The preview runs on the same European infrastructure where the model was trained ¹. Mistral also announces a European version run entirely by the company itself, without relying on other digital providers and under European law ¹.

If you handle health data, contracts, employee data or customer information, this short chain matters. Fewer steps, fewer parties to list in your record of processing activities, fewer awkward questions from your privacy consultant.

With open weights, in theory you will be able to run the model in-house. In practice, a model with 1 trillion parameters ¹ won't fit on the server in the storage cupboard. For an SMB, the realistic route remains the European API, or an Italian provider that hosts it for you.

What the announcement doesn't say

There are no prices. You don't yet know what it will cost you at full scale, and a preview, by definition, can change.

There is also no stable availability date for production. Mistral writes that the model is still improving, which is good news for quality and bad news for anyone who wants identical results from one month to the next.

What to do on Monday

If today you send no sensitive data to any AI model because you don't trust American servers, this is your chance to run a serious test. Pick a single use case: summarising contracts, extracting data from invoices, replying to recurring emails.

Try it on Mistral Studio with fake or anonymised documents. Compare the answers with those from the tool you already use. Look above all at the Italian and the errors, not the speed.

Before you put in real data, ask Mistral for its contractual terms on data processing and the exact region where the service runs. A promise in a post is not a signed DPA.

If, on the other hand, ChatGPT or Claude work well for you and you don't handle sensitive data, do nothing for now. Wait for prices and some independent tests.

If you want to work out which of your processes make sense to move to a European model, that is the kind of assessment we at AIR often do with companies, starting from a single process.

A competitive European model doesn't oblige you to switch. It gives you a choice you didn't have before.