Mistral released downloadable weights for its Large 4 model on Thursday, Oct. 29, allowing organizations to run the trillion-parameter multimodal system on their own systems instead of relying only on Mistral’s hosted service.
The model, nicknamed “Le Chonk,” has 1.05 trillion parameters in total, with 52 billion used at a time, and can process images as well as text. Mistral first introduced Large 4 in public preview earlier this month. The company says it used feedback and three additional weeks of reward-based training to improve the released version.
Mistral is promoting the model for coding and cybersecurity, among other uses. The company says the model performs better and reasons more efficiently. Mistral also says Large 4 performs better than other models with downloadable weights on certain coding tests.
Those results are the company’s claims, not an independent assessment in the release announcement. Making the weights available gives businesses and public agencies more control over where and how they use the model. This could be especially appealing to organizations that handle sensitive data or security work.
The release also adds a large downloadable model to the competition between models with downloadable weights and AI services that users can access only through a provider.