Mistral AI unveiled Mistral Large 4, which it presents as its largest and most capable model to date. The model, internally known by the codename “Le Chonk,” contains one trillion parameters in total, while using 49 billion active parameters in each operation. The company is making it available as a public preview through Mistral Studio, with plans to release the model’s weights before the end of October.
A Multimodal Model for Enterprise Use
Mistral Large 4 processes text and images, and the company says it was designed for programming, workflows based on intelligent agents, image understanding, finance, law, and cybersecurity. Mistral claims that the model achieves among the highest performance levels in a number of enterprise scenarios compared with models with open weights.
In the Dense 200 image-understanding test, the model scored 42% versus 41% for the GPT-6 Astra model, while reportedly being able to analyze complex documents, diagrams, and technical schematics; find specific elements within large satellite images; and identify the locations of elements in visual data.
Performance Indicators in Security and Programming
Mistral Large 4 achieved 82% in a test that reproduces vulnerabilities in real software and then targets their remediation. It also completed 93% of the 40 tasks in the Cybench test. The company says its security applications include malware analysis, vulnerability prioritization, and the creation of detection rules.
In programming tests, the model scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4, and 49.8% on the Coding Agent Index. According to Mistral, it outperformed DeepSeek V4 Pro and Qwen3.8 Max on the overall index, while ranking second behind Claude Opus 5 in a blind evaluation conducted by professional developers.
Agents and European Infrastructure
The model also focuses on executing long-running tasks through the use of multiple tools and applications. In AutomationBench, which covers 657 enterprise workflows across Gmail, Google Sheets, Slack, and Salesforce, it scored 59.9%. It also recorded 1,393 Elo points in the AA-Briefcase test for producing spreadsheets, presentations, and PDF files.
The model was trained from scratch in Mistral’s European data centers using 3,800 Nvidia Grace Blackwell GPUs. The company says it supports more than 160 languages, including all official languages of the European Union. It will also be hosted in Europe through infrastructure operated by Mistral and subject to European legislation, with the goal of enabling enterprises to run the weights in their own clouds or on-premises infrastructure once they are released.
What Changes in Practice?
Mistral Large 4 combines a large multimodal model, agent capabilities, and an anticipated open-weights option, potentially giving enterprises a path to reduce their reliance on external application programming interfaces and gain greater control over data and usage policies. However, details of the architecture, training methods, and additional test results have not yet been published, and most of the performance indicators cited are based on the company’s own data; therefore, independent comparison will have to wait for the weights and detailed information.
The API is priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says the model represents the first major model in the roadmap following a €3 billion Series D funding round, and that it will also serve as the foundation for sector-specific and optimized models it will develop later.