Mistral Large 4: Europe finally has a frontier model of its own

Mistral AI has just launched a public preview of Mistral Large 4, “le Chonk” for friends. It is the biggest model Mistral has ever built, and for those of us who build software in Europe, it's the most important AI release of the year.

What was announced

• 1 trillion parameters, a mixture-of-experts with 52 billion active per token, natively multimodal (text and images in).
• According to Mistral, the best open-weight model developed in the US or Europe, competitive with the strongest open models worldwide, and state of the art among open models on cybersecurity, finance and law. On visual grounding it even edges past closed frontier models (42 % vs 41 % for GPT-6 Astra on Dense 200).
• Strong numbers elsewhere too: 93 % on Cybench, 61.7 % on DeepSWE v1.1, 59.9 % on AutomationBench.
• Trained on a multilingual mix of more than 160 languages, including every official EU language. Yes, Greek and German too.
• Available today as a preview API on Mistral Studio, at $1.36 per million input tokens and $4.18 per million output tokens. Open weights by the end of October.

Forged in Europe, end to end

This is the part that matters most to me. Mistral Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe, and the preview is served from that same infrastructure. Mistral will also run a European deployment end to end, independently of other digital service providers and under European law.

It's also the first milestone funded by Mistral's €3 billion Series D, the largest equity round ever raised by a European technology company. Europe isn't just regulating AI any more. It's building it.

Why sovereign AI matters

I build websites, shops and AI tools for European businesses. In almost every project the same questions come up: where does our data go, who can read it, and what happens if the provider changes the rules?

With the big US platforms, the honest answer is complicated. Even when the servers are in Frankfurt, the company behind them is American, and US law such as the CLOUD Act can reach data held by US companies wherever it is stored. Terms, prices and access can change overnight, decided somewhere else. Mistral makes a sharp point about cybersecurity: a model that refuses legitimate vulnerability research, or disappears mid-incident, is itself a security risk.

Sovereign AI means the opposite: a European model, trained in Europe, run by a European company under European law, and, once the weights are out, a model you can run on your own servers, under your own policies.

GDPR, without the gymnastics

For GDPR, this changes the conversation. Mistral is a European company that answers to the GDPR directly, and its European deployment keeps processing inside the EU. That takes away the hardest question in most data-protection reviews: transfers to third countries, the whole Schrems II headache.

And with open weights you can go one step further. Self-host the model, on-premise or in a European cloud of your choice, and personal data never leaves your infrastructure at all. For healthcare, law firms, accounting, the public sector, or any shop with a customer database, that's the difference between “we'll need a long DPIA” and “we can start next week”.

A fair word of caution: no model makes you compliant by itself. You still need a data processing agreement, a legal basis and sensible data minimisation. But starting from a European, self-hostable model makes all of that much simpler.

The honest caveats

• It's a preview. Mistral says the reinforcement-learning run is still going and the model will keep improving.
• Most benchmarks are Mistral's own, though several come from independent evaluators (Artificial Analysis, vals.ai, Surge AI). Wait for the community to test it.
• Self-hosting a 1T model is not a laptop job. Only 52B parameters are active per token, but you still need serious multi-GPU hardware to hold all the weights. Most teams will start with the API.

Try it

If you want to see what le Chonk can do, these all open in a new tab:

• Try Mistral Large 4 in Mistral Studio: the preview API is live today.
• Model documentation: model ID, limits and how to call it.
• The official announcement: every benchmark, the demos and the roadmap.

I'll be testing it on real client work over the coming weeks, especially the European deployment and, at the end of the month, the open weights. If you're a European business wondering whether you can use AI without sending your data across the Atlantic, get in touch. That question finally has a good answer.