Mistral Large 4 puts Europe back in the open-model race

Europe has a frontier-class open model again. Almost.

Mistral launched a public preview of Mistral Large 4 on Tuesday, a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 49 billion active. The company calls it ML4, and, officially, "le Chonk." The preview API is live on Mistral Studio today. Mistral says it will release the weights by the end of the month.

The specs, as Mistral tells it

Mistral says it trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe. It supports more than 160 languages, including every official EU language, the company says.

The model's documentation lists 1.05 trillion total parameters, a 1 million token context window and a 1.6 billion parameter vision encoder, with the model marked as a public preview.

On Mistral's own numbers, ML4 scores 61.7% on DeepSWE v1.1 for coding and 93% on Cybench, a set of 40 security competition exercises. On the Dense 200 visual benchmark it reports 42%, against 41% for OpenAI's GPT-6 Astra. Mistral says red-teaming with cybersecurity leaders and state authorities is underway.

Those are vendor benchmarks. Treat them as a claim until independent evaluators run the weights.

Open weights are the launch

Yes, another big model with a big benchmark table. What makes this one matter is what happens at the end of October.

Mistral says ML4 is competitive with the strongest open models anywhere and well ahead of any open-weight model from the US or Europe. If the weights land as promised, teams that cannot send data to a US or Chinese API get a large model they can host themselves, trained in a European lab's own data centers.

That is a procurement story as much as a model story. If you buy AI for a European bank, law firm or agency, your choices have been a closed model you cannot inspect or an open one from a lab outside your jurisdiction. ML4 is aimed squarely at that gap.

What you can do today

You can call the preview through Mistral Studio now. Try it on the work Mistral is pitching: code, agent workflows, documents with images, and finance or legal tasks.

Do not move production to a preview. Run your own evals on your own data, keep your current model as the default, and note what it costs you per task, not per token.

If you plan to self-host, start sizing hardware now. A model with a trillion parameters needs serious memory even when only 49 billion are active per token.

What to watch

The date that matters is the weights drop Mistral has promised for the end of October. Then watch the license terms, the independent benchmark runs, and whether the cloud providers put ML4 on their menus.

Preview today. Verdict when the weights ship.