The French company Mistral AI has presented its new flagship model Mistral Large 4, known by the unofficial nickname Le Chonk. It is an open model with a billion parameters that is already available for weight download.
The model belongs to the open-weight category and demonstrates the best results among similar developments in coding and defensive cybersecurity tasks. According to Mistral, it outperforms Chinese models such as Kimi and DeepSeek in these areas.
Le Chonk is a Mixture-of-Experts architecture with roughly 1 trillion total parameters and 49–52 billion active ones. It supports multimodal input of text and images, as well as a context of up to a million tokens.
Training was carried out from scratch on European infrastructure — around 3800–4000 Nvidia Grace Blackwell GPUs in Mistral data centers. This underscores the company's emphasis on sovereign AI, where data remains under the users' control.
The model's weights were released at the end of October 2026 after a period of public preview. During that time, the developers gathered feedback and further trained the system, improving its performance in agentic coding and security tasks.
On benchmarks such as the Artificial Analysis Coding Agent Index, Le Chonk shows 48 %, while using 47 % fewer tokens than competitors. In cybersecurity, the model handles the detection and fixing of vulnerabilities without unnecessary refusals.
The model is available through the Mistral API, Hugging Face and partner platforms. The license permits free use and modification, but with a restriction on commercial hosting as a service without a separate agreement.
How will this affect the balance of power in open AI? The European development offers an alternative to closed American and Chinese systems, giving organizations more control over data and infrastructure.
The release highlights the growing role of open-weight models in areas where security and customization matter. Mistral positions Le Chonk as a tool for developers, companies and government agencies striving for independence in AI.
