Mistral unveils Large 4, a one trillion-parameter multimodal model that is in public preview for now
In 30 seconds
Large 4, which Mistral itself nicknames le Chonk, is multimodal and has a trillion parameters, with 49 billion active. For now it can only be tried through the API as a public preview, and downloadable weights are due before the month ends. It was trained from scratch in Europe, in the company's own data centers, on 3,800 NVIDIA GPUs.
Mistral has unveiled Mistral Large 4, its largest model to date, which the company itself calls "le Chonk".
It is a multimodal model with one trillion parameters, 49 billion of which are active. For now it is in public preview and can be tested via API. The open weights (the downloadable model) will arrive at the end of this month. The company is also running security tests on the model in real-world settings with cybersecurity leaders, vetted partners, and state authorities.
According to Mistral, it already competes with the strongest open models in the world and significantly outperforms any open-weight model developed in the US or Europe. The company trained it from scratch on 3,800 NVIDIA GPUs in its own data centers in Europe, and a significant share of the training data was multilingual, spanning more than 160 languages.
Why it matters · analysis and opinion
A model this size with open weights widens the options for anyone who does not want to depend on a provider from outside Europe: it will be possible to download it and run it on your own servers, which is key when data cannot leave the company. The comparisons with other models come from Mistral itself, so it is worth waiting for independent tests. In our view, the useful step for a small business today is not to install it, since a trillion parameters demand a lot of hardware, but to try it through the API on real tasks and watch which downloadable versions arrive at the end of the month, and under what license.
Official source: Mistral AI · Written with the help of AI: how we make the news
