Mistral Large 4: France's largest AI model in preview

What's new

On October 6, 2026, the French AI company Mistral opened a public preview of Mistral Large 4, called ML4. Internally, it also goes by the nickname "Le Chonk." It is the company's largest and most capable model to date.

The model has 1,000 billion parameters (the "knobs" an AI model has learned to turn). Only 49 billion of them are in use at a time, because the model is built as a so-called Mixture of Experts: a kind of team of specialists where only the relevant ones take on the task. That makes it faster and cheaper to run. It understands both text and images and can switch between quick answers and longer reasoning. It was trained on around 3,800 Nvidia chips in Mistral's European data centers.

The preview can be tested now in Mistral Studio. According to Mistral, the weights, that is, the model's internals, will be released at the end of October. Euronews mentions October 27.

What's clever

Mistral points to three things in particular. First, language: the model speaks more than 160 languages, including all of the EU's official languages. Second, security: on IT security tests, it scores 93 percent on Cybench and 82 percent on reproducing vulnerabilities. Mistral's CEO Arthur Mensch says that in some areas, including cyber, the model beats Chinese models. Third, it is built in Europe.

On coding, Mistral has measured 61.7 percent on DeepSWE v1.1 and 28.3 percent on Terminal-Bench 4.0. On image understanding (pointing to the right spot in an image), it gets 42 percent against 41 percent for OpenAI's GPT-6-Astra. Mistral also says the model beats GPT-6-Astra on legal and financial tasks. Keep in mind that the figures come from Mistral itself and have not yet been independently verified. The length of what the model can keep track of at once (context length) has not been disclosed.

Cheaper or better?

The API price, which developers and companies pay, is 1.36 dollars per million tokens (pieces of text) in and 4.18 dollars out. By comparison, its predecessor Large 3 cost around 0.50 dollars in and 1.50 dollars out, so Large 4 is roughly twice as expensive. On the other hand, it is a far larger model. We have not found prices for competitors' models, so we cannot say whether it is cheaper than, for example, GPT-6-Astra.

The price for ordinary users in Mistral's chat app Le Chat has not been disclosed. Whether the model will be free for individuals has not been settled either.

Mistral calls the model the strongest open model outside China. Open means that companies can download and run it themselves. The license has not yet been disclosed.

What it's good for

The model is aimed especially at companies that work with code, IT security, law, finance and manufacturing, and that want an AI that can solve tasks on its own (agents). It can also read images and documents.

For Danish and European companies, the point is control: data can be kept in Europe, and in principle you can run the model on your own servers once the weights are released. That can be important for public authorities and regulated industries. Mistral recently raised 3 billion euros at a valuation above 21 billion euros and is well positioned as Europe's answer to American and Chinese AI models. For now it is a preview, so the final results may change.