Mistral Opens a Preview of Large 4, a 1 Trillion-Parameter Model It Calls 'Le Chonk'

The Paris-based company says the model leads open-weight rivals built in the U.S. and Europe on coding and cybersecurity tests, and it plans to release the weights by the end of October.

PARIS - Mistral AI on Tuesday released a public preview of Mistral Large 4, its largest model to date, and gave it a nickname to match. “Unofficially ML4, very officially: le Chonk,” the company wrote in its announcement. Developers can try the preview through an API on Mistral Studio now, and Mistral said it will release the model’s weights by the end of the month.

ML4 has 1 trillion parameters, with 49 billion active at a time, and is natively multimodal, meaning it was built to handle images as well as text. Mistral said it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe, and that the preview runs on the same infrastructure. More than 160 languages were represented in the training data, including every official language of the European Union, the company said.

Mistral’s results for ML4 on the Artificial Analysis Cyber Index Chart: Mistral AI

The company’s central claim is that ML4 is the strongest open-weight model developed outside China. On the Artificial Analysis Cyber Index, an independent evaluation of how well models find and fix security flaws, Mistral said ML4 ranks among the top five models globally. On one of the index’s tests, which asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scored 82 percent, which Mistral said is the highest of any model. It also said several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that test because they refuse the task.

Mistral’s DeepSWE 1.1 coding results for ML4 Chart: Mistral AI

That is a pointed argument from a company selling to security teams. Mistral said refusals by closed models can block legitimate vulnerability research, and that open weights let organizations run the model on their own hardware under their own policies. Before the weights ship, the company said, it is red-teaming ML4 with cybersecurity leaders, vetted partners and state authorities, who are using a version with reduced moderation and expanded cyber capabilities.

On coding, Mistral reported scores of 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 28.3 percent on Terminal-Bench 4. In a blind human evaluation run with Surge AI, professional annotators ranked ML4 second of five models on coding quality, behind Anthropic’s Claude Opus 5, according to the company.

Mistral’s B3 Agent Security Benchmark results Chart: Mistral AI

Mistral also published safety figures. It said ML4 resisted 93.3 percent of attacks on Lakera’s public B3 AI Security Benchmark, and that its refusal rate on malicious cybersecurity prompts is higher than that of other open models it tested.

Training reward trends from Mistral’s reinforcement learning run Chart: Mistral AI

All of these results come from Mistral’s own announcement, and the model is still a preview. Mistral said its reinforcement learning run is still in progress and that it expects “large and rapid improvements” as training continues. It said ML4 is the first milestone funded by its 3 billion euro Series D round, and that it will serve as the base for a new set of specialized Mistral models.

The Verge noted that Mistral is positioning ML4 as particularly strong at coding, cybersecurity and tasks in finance, engineering and manufacturing. The real test will come at the end of October, when outside researchers can download the weights and check the numbers for themselves.