Mistral, the French artificial-intelligence company, has put out a public preview of its largest model yet, a trillion-parameter system it says can match or beat the leading Chinese open-weight models in several fields.
The model is Mistral Large 4, nicknamed Le Chonk, and the preview went live on October 6.
It has one trillion parameters, of which 49 billion are active at a time, and is natively multimodal. It takes in up to one million tokens at a time and the company says it is fluent in more than 160 languages, including every official language of the European Union.
Mistral says it was trained from scratch on 3,800 Nvidia Grace Blackwell chips in the company’s own European data centres, and that the preview runs on the same machines.
The weights are due at the end of the month. Reuters reported a release date of October 27.
Until then, cybersecurity specialists and state authorities will be given a version with fewer safety limits, so that its capabilities can be tested.
Guillaume Lample, the company’s co-founder and chief scientist, told ZDNet: “The cyber defence capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.”
The company’s own note makes the same point from the other side: Firms and states, it said, should not have to rely on a closed-model vendor that could “arbitrarily turn off their cyber defence capabilities”.
On a test of reproducing software vulnerabilities, Mistral said its model scored 82 per cent while two leading closed systems it named — Anthropic’s Claude Opus 5.5 and GPT-6 Astra — scored close to zero because they refused the task outright.
Arthur Mensch, the chief executive, told a conference in Abu Dhabi that the model was ahead of Chinese systems “on certain aspects, including cyber”.
“The narrative that Europe cannot compete is something that is not true,” he said.
Pierre Stock, vice-president of science, told Reuters it ranked among the best open-weight options. He also said the system had tried to go beyond its testing environment, and that this had been expected.
On its own charts Mistral puts a combined coding-agent score at 49.8 per cent, ahead of DeepSeek V4 Pro and Qwen 3.8 Max, and claims state-of-the-art results among open models in cybersecurity, finance and law.
In visual grounding it says it beats even closed frontier models.
An independent reading is less flattering: Artificial Analysis scored the preview at 38.4 on its intelligence index, against nine for Mistral Large 3, but placed it eighth among open-weight models.
Seven Chinese systems rank above it, led by Xiaomi’s MiMo-V2.6-Pro on 46.3 and Z.ai’s GLM-5.3 on 44.8. Claude Opus 5.5, which is closed, scores 57.6.
The launch is the first new Mistral model since May.
In August the three-year-old company began offering customers third-party systems, starting with GLM 5.3 from the Beijing firm Z.ai.
It raised €3 billion in its latest round and sells itself as a European alternative to American closed models and to cheaper Chinese open-weight ones. The round, led by Samsung Electronics, valued the company at more than €21 billion and was presented by Mistral as the biggest equity raise by a European technology firm.
A European deployment, run by Mistral alone and under European law, is planned alongside versions in other regions. A second data centre is being built in Sweden, and the company has said it is aiming for 200 megawatts of European capacity by the end of 2027.
What happens when the government can see almost everything? 🔍👀
AI + financial data + automated enforcement could fundamentally change the relationship between citizens and the state.@robertamsterdam joins Brussels Signal to discuss Spain, surveillance, taxation, de-banking,… pic.twitter.com/AarVD8FF1X
— Brussels Signal (@brusselssignal) September 10, 2026