Introducing Mistral Large 4
Mistral ● Covered by 4 sources
Mistral previewed Large 4, a 1T-parameter open-weight model built in Europe. It’s aimed at coding, cyber, and multimodal work, with weights due by month-end.
Based on reporting by Mistral — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral has opened a public preview of Mistral Large 4, the company’s biggest model yet. The nickname is le Chonk, which feels about right for a system with 1 trillion parameters and 49 billion active ones. You can try it now through Mistral Studio, and the weights are set to arrive by the end of the month.
The pitch is straightforward: this is Mistral trying to claim the high end of open-weight AI. The company says ML4 is competitive with the strongest open-source models overall, beats any open-weight model built in the US or Europe, and is state of the art on enterprise-heavy work like cybersecurity, finance, and law. In some visual grounding tests, Mistral says it even edges past frontier closed models.
Cybersecurity gets the loudest emphasis, and for good reason. On the Artificial Analysis Cyber Index, ML4 sits among the top five models globally, and on one test for reproducing and patching a real vulnerability it scores 82%, which Mistral says is the best result of any model. It also solves 93% of Cybench’s 40 challenges. Closed models such as Claude Opus 5.5 and GPT-6 Astra reportedly score near zero on that same test because they refuse to do the task. That’s the whole argument in one messy, very current AI sentence: if defenders need to prove a flaw exists, a model that won’t touch the work is not much help.
Mistral is also leaning hard on the fact that ML4 was trained from scratch in its own European datacenters on 3,800 NVIDIA Grace Blackwell GPUs. The public preview runs on that same infrastructure, and the company says a European deployment will be operated end to end under European law, independent of other digital service providers. The multilingual training data included more than 160 languages, among them every official EU language.
Beyond cyber, the benchmark story is broad. Mistral says ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. It scores 59.9% on AutomationBench, which covers 657 business workflows across tools like Gmail, Google Sheets, Slack, and Salesforce, and reaches 1,393 Elo on AA-Briefcase. On visual grounding, Mistral cites a 42% score on Dense 200 versus 41% for GPT-6 Astra. The company also says the model is strong on science, math, legal work, and finance, and that it will underpin a new generation of specialized Mistral models.
My take — AI-written commentary, not fact-checked reporting
This is the kind of launch that makes open weights look serious instead of sentimental. Europe keeps saying it wants sovereignty; Mistral is at least building something that sounds like a real answer, not a policy deck with a GPU sticker on it. The fun part is that the strongest argument for open models here is also the least glamorous one: letting defenders do their jobs without asking a chatbot for permission.
Read more about this at: Mistral