Mistral unveils a top European model you will soon be able to run yourself
On 6 October Mistral unveiled Mistral Large 4, trained in its own European data centres. At the end of October it plans to release the model itself, so you can run it on a server of your own.
What happened
On 6 October the French company Mistral unveiled Mistral Large 4, its largest model so far. It was built in Mistral’s own European data centres, and the company plans to release it at the end of October, so you can run it on a server of your own.
Large 4 has a trillion parameters, the numbers in which a model stores what it has learned. For each piece of text it uses 52 billion of them; early reports said 49 billion. The model understands text and images, such as documents, charts, technical drawings and satellite imagery, and works in more than 160 languages, including every official EU language. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell chips.
For now there is a preview on Mistral’s own platform, at $1.36 per million tokens (chunks of text a few letters long) going in and $4.18 per million coming out. It runs on the same European servers as the training. Later the model will be offered in more regions, including a European setup that Mistral runs entirely itself. It is not finished yet: the final training run is still going. Details of its design and more test results will follow with the weights, the model itself. Mistral has not yet said which licence they will carry.
Mistral calls Large 4 one of the world’s strongest models for cybersecurity. In a blind test, in which reviewers compared code from five models side by side, it came second, behind Claude Opus 5. All of these figures come from Mistral itself. Until the weights are public, security firms, partners and government bodies are testing the model with looser filters. Three days earlier the German company Aleph Alpha released Kolibri: a much smaller open model for German and English, trained in Germany and Finland.
“A European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law.”
Mistral on where Mistral Large 4 runs, 6 October 2026
The background
Why such a big model only uses a slice per word
Large 4 is what is known as a mixture of experts. The model is made up of many sub-networks, the experts, and for each piece of text it picks a few of them. That way it only computes about a twentieth of the model, which makes answers faster and cheaper. But which experts are needed changes from word to word, so all of them have to sit ready in the chips’ memory.
That memory is the real hurdle. According to a calculation by Yotta Labs, the weights take up some 525 GB even when heavily compressed, and about 1 TB at 8 bits per number. The starting point is a server with eight Nvidia H200 chips. A fast 32 GB graphics card holds about 6% of it. It remains an estimate, as the weights are not out yet.
Large 3 was smaller and free to use
The previous version, Large 3, came out in December 2025: 675 billion parameters, 41 billion of them active at a time, trained on 3,000 Nvidia H200 chips. Mistral released it under Apache 2.0, a licence that lets you use a model freely, adapt it and make money with it. For Large 4 that is still open. Reports say Mistral will use a licence of its own, but the company has not confirmed this.
Where the money and chips come from
Mistral has been building its own computing power for some time. In June 2025 it announced Mistral Compute with Nvidia, a European AI cloud with 18,000 Grace Blackwell chips. In 2025 the Dutch chip-equipment maker ASML led a €1.7 billion investment. On 8 September 2026 Mistral raised €3 billion in a round led by Samsung, at a valuation above €21 billion. According to Euronews it is spending €4 billion on data centres in France and Sweden, and more than 125 large companies use its technology. Microsoft and Nvidia have been shareholders for a while.
Our analysis
The open weights are the news, not the European label
It is nice that Large 4 was trained in Europe, but what matters for you is something else: the weights will be public. Then anyone can run the model, including on a server in your own office. You are not tied to one provider, and a model you have downloaded cannot be taken away from you or quietly changed. That independence does depend on a licence nobody has seen yet.
“Run it yourself” sounds simpler than it is. This model needs at least a server with eight heavy-duty AI chips, which is not a realistic purchase for most companies. For them, a smaller open model on a server of their own is the more obvious choice. And the test results come from Mistral itself. The most striking comparison, on reproducing and patching security flaws, mainly shows that Claude and GPT refuse that task. Run such a model yourself and you set the filters yourself, and you carry the responsibility for them too.
Our expectation: open models at this level will become an ordinary option for sensitive data, and the smaller ones run on a server that fits in your own office. Test them on your own work before you choose.
What stands out
- The licence is still unknown Large 3 came under Apache 2.0 and could be used freely, commercially too. Mistral has not yet named a licence for Large 4.
- Everything has to sit ready Large 4 computes only a twentieth of the model per word, but the whole model has to be in memory: some 525 GB, even compressed.
- The best score leans on refusals At reproducing and patching a real security flaw, Large 4 scores 82%. According to Mistral, Claude Opus 5.5 and GPT-6 Astra score close to zero because they refuse the task.
- Test it now and you test a draft The final training run behind the preview is still going. The model released at the end of October may perform differently.
What it means for you
If you work with sensitive data, a strong AI model that you keep in Europe, or on your own server, is coming within reach.
- Your data stays in-house Run an open model yourself and your data does not go to someone else’s AI service. That means fewer questions about privacy and about who else can see it.
- Big is not always needed The bigger the model, the heavier the server it needs. For many tasks a smaller model is enough.
- European says nothing about quality Test a model on your own work before you choose. Trending Topics, for one, finds that Kolibri does not yet match the best open models.
Where we come in · Advisory
We help you find the model that fits your work
In a Discovery or Roadmap we look at which work AI can take on and which model suits it: a cloud service or an open model on an AI server of your own. We supply that server too: an Apple Mac Studio in your office, cut off from the internet and frugal with energy, running models such as Mistral, Gemma and gpt-oss.
Not sure your data can go to an AI in the cloud?
Tell us what work it is about. We’ll think it through with you and always reply within one working day.
