French AI lab Mistral has put its newest flagship, Mistral Large 4, into public preview. The model is natively multimodal and has roughly one trillion parameters in total, though only about 49 billion of them are active when it generates an answer. Developers can try it now through Mistral Studio, and the company says the full weights will be published before the end of October 2026.
Built and trained in Europe
Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs located in its own European datacentres, and that the preview is served from the same infrastructure. That detail matters for customers who want AI workloads to stay under European jurisdiction. The training data covers more than 160 languages, including every official language of the European Union. Internally the model has been nicknamed “le Chonk”, a joke that started life as a meme.
A strong focus on cybersecurity
Much of the announcement leans on security results. Mistral reports that Large 4 places in the top five on the Artificial Analysis Cyber Index, a benchmark that measures how well a model can find and repair flaws in real software. On a task that asks a model to reproduce a genuine vulnerability in open-source code and then patch it, the company claims a score of 82 percent, which it describes as the best result recorded. It also says the model solves 93 percent of the 40 challenges in Cybench.
Mistral is also taking an unusual position on safety controls. It argues that blanket refusals from AI providers can get in the way of legitimate vulnerability research and incident response. To address this, vetted cybersecurity partners and government bodies are testing a version of the model with lighter moderation and extra cyber capabilities. When the weights are released, organisations will be able to run it in a private cloud or on their own servers and apply their own policies.
One caution: the company also reports that the model withstood 93.3 percent of attacks on a public prompt-injection benchmark from Lakera. That is a benchmark figure, not a guarantee that it will resist nearly every real-world attack.
Coding agents and business workflows
For software engineering, Mistral cites 61.7 percent on DeepSWE v1.1 and 28.3 percent on Terminal-Bench 4, with a combined coding-agent index score of 49.8 percent. In a blind test run with Surge AI, professional annotators ranked Large 4 second out of five models for coding output, scoring 3.74 out of 5. The top spot went to Claude Opus 5 with 4.22. On AutomationBench, which covers 657 workflows across tools such as Gmail, Google Sheets, Slack and Salesforce, the model scored 59.9 percent. Mistral also showed it reading technical drawings, PDFs and satellite imagery.
All of these numbers come from Mistral or from benchmarks it selected, so independent testing after the open-weight release will give a clearer picture.
How it was trained
The model uses the same training and reinforcement-learning environment that Mistral offers customers through its Forge product. Its training tasks blend chat, science problems, safety alignment, factual accuracy and long tool-using jobs, with answers checked by reward models, unit tests, LLM judges and static analysis. At a scale of 3,000 GPUs, the company says one run produces about 33 billion tokens a day, around 16 billion of which are usable for training. That reinforcement-learning run is still going.
What to watch next
The key date is the end of October, when Mistral has promised the weights, architecture details, more benchmarks and its post-training method. If the release arrives on time, Large 4 would be one of the largest open-weight models available, and a notable European alternative to US and Chinese systems.
Source: reporting by AI News (artificialintelligence-news.com), 6 October 2026, and Mistral AI’s announcement. Benchmark figures are as reported by Mistral.