Mistral Large 4: trained model from scratch in 3,800 GPUs in Europe arrives with open weights

Mistral launched Mistral Large 4, its most ambitious model so far, trained from scratch in 3,800 NVIDIA Grace Blackwell GPUs within own data centers in Europe. Weights are open, and the company presents the movement as the biggest step of a long-term investment in infrastructure, research and product. In practice, the announcement places the French company as a European alternative for organizations that want cutting edge performance without giving up control over where and how the model runs.

What did Mistral announce with Mistral Large 4?

According to the official announcement, Mistral Large 4 was trained from scratch — it is not a fine fit on a ready base — in 3,800 NVIDIA GPUs Grace Blackwell installed in Mistral's own data centers in Europe. The company describes the result as cutting edge performance in critical verticals, with open weights made available to customers. The communiqué frames the launch as part of a long-term strategy that combines infrastructure, research and product, and not as an isolated delivery.

Why does training Mistral Large 4 in its own infrastructure change the equation?

Training a frontier model requires concentrated computational capacity for weeks or months. By using Grace Blackwell GPUs — NVIDIA architecture that combines Grace CPU and Blackwell GPU for large-scale training loads — within own data centers, Mistral reduces third-party cloud dependency and maintains control over the full cycle from training to inference.

This directly talks to the positioning that the company has been building. We have already reported here on the portal Mistral's bet on regional inference and open models for European sovereignty, and Mistral Large 4 takes this logic a step back in the chain: now it is not only where the model runs, but where it is trained.

Open Weights of Mistral Large 4: what does this release in practice?

Open weights mean organizations can download model parameters, run it on their own infrastructure, and adapt it to specific domains without relying on an external API. For regulated sectors — health, finance, defense, government — this is often the difference between adopting or not a border model.

Mistral was already signaling this direction in the corporate offer, with open weight models for companies and governments. Mistral Large 4 expands the capacity ceiling of this strategy.

What does Mistral Large 4 change for those who develop in Brazil?

For Brazilian teams, the combination of open weights and European training opens up three concrete paths. The first is architectural: it is possible to run the on-premise model or national cloud, keeping sensitive data within specific jurisdictions — a relevant point for projects that need to dialogue with LGPD. The second is economical: open weights eliminate the cost per token in high volume scenarios, changing it for fixed cost of infrastructure. The third is product: teams can adjust the model to Portuguese and local domains without depending on supplier roadmap.

It is worth noting that Mistral is not the only one to bet on a massive training scale — NVIDIA has already announced historical AI infrastructure projects with other partners, such as the megaproject with OpenAI. The difference here is the combination of scale, open weights and European jurisdiction.

What the Mistral Large 4 announcement hasn't solved yet

The statement highlights vertical infrastructure and performance, but does not replace independent evaluation: reproducible benchmarks, detailed weight licensing, context limits and actual cost of production operation are what will determine large-scale adoption. As long as these numbers do not circulate outside Mistral's own material, architectural decisions based solely on the announcement are premature.

Frequently Asked Questions about Mistral Large 4

How many GPUs have been used to train Mistral Large 4?

There were 3,800 NVIDIA GPUs Grace Blackwell, according to Mistral, in the company's own data centers in Europe.

Does Mistral Large 4 have open weights?

Yeah. Mistral states that the model is distributed with open weights, which allows execution and adaptation in own infrastructure.

Where was the model trained?

The training occurred from zero in data centers of Mistral itself located in Europe, without using a pre-trained third-party base.