Mistral Large 4: European giant model with 1 trillion parameters

What's new

French AI company Mistral has unveiled Mistral Large 4, called ML4 and nicknamed "Le Chonk." The model was shown off on October 6, 2026, and is the company's largest to date. It is multimodal, so it can work with text, documents, charts and images, and it supports more than 160 languages, including all of the EU's official languages.

The model has 1 trillion parameters in total. Parameters are the settings an AI model learns during training, and the more there are, the more it can hold. But ML4 uses a so-called mixture-of-experts architecture: Only part of the model is active at a time. Mistral states 52 billion active parameters, while one source cites 49 billion. That typically keeps costs down.

ML4 is a so-called open-weight model. That means the model's "weights" are released, so others can download and run it themselves. That will only happen later, though. Mistral wants to test safety first and work with trusted partners and governments. According to Mistral, the weights will arrive before the end of October, and other outlets cite October 27. Until then, the model can be tried as a public preview via the API in Mistral Studio.

What's smart

Mistral trained the model from scratch on Nvidia Grace Blackwell chips in European data centers. That involves 3,800 chips according to Mistral's own site and roughly 4,000 according to TechCrunch. Mistral's VP Science, Pierre Stock, says that is two to three times fewer chips than at the Chinese competitors and markedly fewer than at the closed models.

The other major selling point is European sovereignty. The model can be run in a European deployment that Mistral operates itself from end to end, independent of other providers. For companies and government agencies that don't want to send data out of Europe, that is an important point.

Then there is cybersecurity. According to ITPro, ML4 solved 93 percent of the tasks in the Cybench test, while Claude Opus 5.5 and GPT-6 Astra scored near zero, because the closed models refuse that kind of task for safety reasons. Mistral itself says the model ranks in the top 5 on the Artificial Analysis Cyber Index and handles 82 percent of the tasks involving reproducing vulnerabilities.

Cheaper or better?

The API price is $1.36 per million input tokens and $4.18 per million output tokens. Tokens are the small chunks of text the model reads and writes. No monthly price has been given for regular users, and the sources do not make clear how the pricing compares with competitors.

On quality, the picture is mixed, and the figures come mostly from Mistral itself. The company claims that ML4 beats GPT-6-Astra on finance and legal tasks in third-party tests, and that it is the strongest open-weight model developed outside China. On visual grounding, the result is 42 versus 41 percent, a razor-thin lead. Full comparisons with Chinese models such as DeepSeek, Kimi and Qwen are not yet available, according to TechCrunch.

On coding, Mistral is candid: The model lags behind the best. The company reports 61.7 percent on DeepSWE v1.1 and 59.4 percent on SWE-Atlas-QnA.

What it's good for

Mistral itself points to cybersecurity, finance, manufacturing and chip design. Here, a model that doesn't refuse security tasks and can be run under European control may make sense for companies with sensitive data.

For pure programming, it is a less obvious choice. The license terms for the upcoming weights are not yet known, and the sources do not say whether the offering applies separately in Denmark. The API is open, however, via Mistral Studio. Mistral positions itself between closed American models and open Chinese alternatives, and the company's investors include ASML and Samsung.