Technology and Scienceupdated 2h ago2 min read
Mistral Launches Mistral Large 4 “Le Chonk” Open-Weight Model To Compete With U.S. And China
Mistral launches Mistral Large 4, a 1-trillion-parameter multimodal model. The model is nicknamed 'Le Chonk' and targets US and Chinese rivals.
14 outlets told this the same way.
ML4, open weights, Oct 27
Mistral AI unveiled Mistral Large 4, nicknamed "Le Chonk," as an open-weight, multimodal model designed to process text and images. Mistral said ML4 would be available in a pre-release on Tuesday via its API, and it planned to publish the weights on October 27 in an open-weight format. Mistral described ML4 as having 1.05 trillion parameters in total, with 49.0 billion activated during each inference, and it said the model uses a mixture of experts architecture.
Mistral said it trained ML4 from scratch for two months on 4,000 Nvidia Grace Blackwell GPUs hosted in its European data centers, and it positioned the release as a way to compete with American and Chinese models. Mistral also set pricing for its model, including 0.68 dollars per million input tokens, 0.07 dollars per million cached input tokens, and 2.09 dollars per million output tokens.
Cyberdefense and agentic aims
Mistral highlighted cybersecurity as a core application, saying companies would be able to develop their own defense tools against cyberattacks without relying entirely on closed models. Guillaume Lample said, "most companies are vulnerable and that closed models sometimes raise access issues," and he added that ML4 is intended to help defense against malicious actors. Arthur Mensch said during a conference in Abu Dhabi that ML4 is “superior to Chinese models in some aspects,” particularly in cybersecurity.
Mistral also framed the model as agentic, saying ML4 is intended to help design 3D objects for industry, synthesize large volumes of documents, assist developers in coding, and identify objects in satellite or aerial images. Mistral’s rollout included a public beta accessible through Mistral’s API in Mistral Studio, while the company said the weights would not be published until October 27.
Funding, compute, and the race
Mistral said its ML4 release follows a fundraising push, with a Series D of €3 billion closed in September and a funding round led by Samsung. Mistral framed the model as trained on 4,000 Nvidia Grace Blackwell GPUs for two months in Europe, and it said it aims to have one gigawatt of computing power by 2030. Mistral’s CEO Arthur Mensch used a public message on X to describe the training and deployment on its own computing infrastructure, saying, "reinforcement learning shows no signs of fatigue," and he said the company ensured ML4 could be deployed with a minimal configuration.
Mistral also said it planned to work with developers, cybersecurity specialists, and public authorities to "more closely assess the capabilities and behavior" of the model under real conditions before weights are published. Mistral positioned the launch as a bid to rejoin the global AI race, with Mistral’s leadership claiming the model ranks "among the best" in its category in terms of performance.