Mistral releases open-weight Le Chonk, its 1-trillion-parameter coding model

Mistral AI released the full weights of Mistral Large 4, an open-weight model it calls Le Chonk, the company said. Weights are the trained internal settings that determine how a model responds, and publishing them lets anyone download, inspect or run the system on their own servers instead of Mistral’s. The model has 1 trillion parameters, the adjustable values a neural network tunes during training, and Mistral describes it as natively multimodal, meaning it was built from the start to handle text and images together. The company trained it entirely on European infrastructure.

Mistral says the updated model is now the strongest open-weight option for coding among publicly available systems. On Cognition’s private FrontierCode 2.0 benchmark, which checks whether a repository maintainer would actually merge the code change an AI agent proposes, the company reports a score of 19.7%, the top result among the models it compared. Mistral also says Large 4 leads open-weight models on Terminal-Bench 4.0, a test of command-line coding tasks, reaching 41%, a 13 percentage point jump from the preview version released earlier this month, while cutting output tokens by 30%. Tokens are the text chunks a model generates and is billed for, so using fewer of them lowers the cost of running a task.

On the Artificial Analysis Coding Agent Index, an independent benchmark, Mistral says Large 4 scored 48%, roughly matching the open-weight model GLM-5.3, while using 47% fewer tokens per task on average. The company also points to blind pairwise tests run by Surge Code, in which it says human graders preferred Large 4’s answers over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking the model behind only Anthropic’s Claude Opus 5.5 among the systems tested.

Built for defenders

Mistral says Large 4 scores 51% on the Artificial Analysis Cyber Index, a benchmark that measures a model’s ability to find vulnerabilities in a codebase, confirm they are real, and patch them without breaking the software. The company says it built guardrails meant to separate legitimate defensive security work from malicious requests, and that the model refused malicious cybersecurity prompts 96.6% of the time on average across three test sets, the highest refusal rate it has measured among open-weight models.

Several cybersecurity vendors are quoted in Mistral’s announcement describing their own testing of the model. Booz Allen Hamilton’s Brad Medairy said internal testing matched the model’s vulnerability-detection results against other frontier cybersecurity models at a lower cost, and that Mistral’s support for self-hosted deployment helps meet client data-residency requirements. Executives from Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, NVIDIA, Rubrik and Zscaler are cited making similar points about open-weight models giving security teams more control over the AI they run. These are claims supplied by Mistral and the companies themselves, not independently verified results.

Mistral also says the European Union Agency for Cybersecurity, ENISA, is testing Large 4 to identify vulnerabilities, audit code and analyze defensive logs. That evaluation, along with any independent benchmarking of Mistral’s own coding and cybersecurity scores, will show whether Le Chonk holds its lead once outside testers get their hands on it.