European AI flag bearer Mistral’s new open weights model is ‘Le Chonk’
Mistral has unveiled Mistral Large 4, a trillion-parameter open weights model it nicknames “Le Chonk”, its largest release to date. The model is designed for multimodal and reasoning tasks, and its European training and deployment are relevant to efforts to build sovereign AI capabilities within Europe.
The model uses a mixture-of-experts architecture with about 49 billion active parameters and was trained on roughly 3,800 Grace Blackwell GPUs using data spanning more than 160 languages. Independent testing places its preview below leading OpenAI and Anthropic models, but ahead of the US open weights model Inkling. Mistral says it performs strongly on cybersecurity tasks; the weights are expected on Hugging Face and other repositories later this month.
- Mistral’s largest open weights model has one trillion parameters.
- Le Chonk was trained on European infrastructure across more than 160 languages.
- Its weights are expected to be released later this month.
New here? Start with this
Mistral is a European artificial intelligence company based in France that develops large language models – sophisticated software trained on vast amounts of text to understand and generate human language. The company has unveiled its largest model to date, called Mistral Large 4, which it has given the nickname "Le Chonk". Unlike many competitors, Mistral releases its models as "open weights", meaning the underlying code is publicly available for others to use and study rather than kept secret.
The launch is significant because Europe is attempting to build its own artificial intelligence capabilities independent of American technology companies that currently dominate the field. Mistral is being positioned as a flag bearer for this effort, developing powerful AI technology that can be trained and deployed within Europe rather than relying on foreign systems. This reflects concerns amongst European policymakers and businesses about technological sovereignty and reducing dependence on American technology.
Independent testing suggests Mistral Large 4 performs below the most advanced models from American companies like OpenAI and Anthropic, though ahead of other comparable systems released as open weights. The company says the model excels at cybersecurity-related tasks and has been trained using thousands of powerful computer processors on data spanning over 160 languages. The underlying weights are expected to be made publicly available later this month through platforms like Hugging Face, an online repository where AI researchers share and access models.