Mistral AI has launched Mistral Large 4, a new large language model with 1,000 billion parameters, of which 49 billion are active during each request. The model, internally nicknamed "Le Chonk," was trained from scratch using 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers over a period of two months. It is currently available in public preview through the Mistral Studio API, and the model's weights will be published on October 27, allowing users to run it on their own infrastructure. The model is designed to handle complex tasks across various domains, including cybersecurity, finance, and industrial production, and Mistral describes it as "state-of-the-art" in several areas. Mistral Large 4 has shown strong performance on multiple benchmarks, including the Legal Agent benchmark from Harvey, where it achieved a 15% score, outperforming some Chinese open-weight models. It also excels in object detection and geospatial analysis, and supports over 160 languages, including all 24 official languages of the European Union. Mistral highlights the model's potential in cybersecurity, stating that it can assist companies in defending against cyberattacks by providing a model that can be used unlimitedly without relying on closed, proprietary models. During the preview period, a less restricted version of the model will be available for cybersecurity experts and state authorities. Mistral's CEO, Arthur Mensch, stated that Mistral Large 4 is "competitive with the strongest open-source models globally" and "significantly surpasses all open-weight models developed in the United States or Europe." The model is part of Mistral's broader strategy to position itself as a leader in open-weight models and to address concerns about digital sovereignty, which refers to the ability of countries to control their own digital infrastructure and data. Mistral raised 3 billion euros in September, which it plans to use to increase its computational capabilities and accelerate research. However, the company faces competition from American and Chinese models, which have significantly more computational resources. Mistral aims to close the gap with these models and has announced plans for a major update within two months. The model's weights will be published on October 27, allowing users to run it on their own infrastructure. Mistral also emphasizes its focus on cybersecurity, noting that some closed models refuse to execute tasks related to vulnerability research, which can be a critical security risk. On the Cyber Index of Artificial Analysis, Mistral Large 4 ranks fifth among 18 models, with a score of 49.5 points. Mistral highlights that its model can handle tasks that closed models refuse, providing a more reliable solution for cybersecurity defense.