Trained on 4,000 Nvidia Grace Blackwell GPUs at European data centers
French AI startup Mistral launched a public preview Tuesday of its open-weight artificial-intelligence model, Mistral Large 4, nicknamed Le Chonk. The company said it will release the model’s trained parameters, or weights, on Oct. 27.
The preview is intended to let developers, cybersecurity experts and state authorities evaluate the model before the formal release.
Mistral said it trained the new model on about 4,000 Nvidia Grace Blackwell graphics processing units deployed at its own data centers in Europe. The company said the model would rank among the top open-weight models globally on aggregate benchmark performance, and described it as the strongest open-weight model developed outside China by a substantial margin.
“The model capabilities will further improve as we scale up our training capacity,” Guillaume Lample, Mistral’s co-founder and chief scientist, said.
Since its early days, the company has advocated for open-weight models — systems whose trained parameters can be downloaded and modified by users. Mistral has said open-weight systems give developers flexibility to tailor models to their needs while protecting their data and intellectual property.
The company also argued that closed models can pose security risks, saying some such systems block responses to useful cyberdefense requests. The preview period is intended, in part, to give security researchers a chance to probe how the new model behaves.
The release positions Mistral against U.S. AI market leaders including ChatGPT maker OpenAI and Anthropic. The company has framed a more capable model as central to its effort to attract more consumers and enterprises and to grow its customer base.
The announcement comes about a month after Mistral completed a funding round led by South Korea’s Samsung Electronics, a transaction the company said lifted its valuation above $24 billion. Mistral has said the additional capital is intended to support further model development.