Mistral AI releases 1 trillion parameter Mistral Large 4 model

French AI lab Mistral on Tuesday unveiled Mistral Large 4, a multimodal model nicknamed "Le Chonk" designed to compete with the top-tier systems from the US and China.

Mistral Large 4 (ML4) features 1 trillion parameters and is currently available in preview for developers, state authorities, and cybersecurity leaders. According to TechCrunch, the model is not yet open-weight, but Mistral plans to release its weights in approximately three weeks following safety testing. Ars Technica reported that the company expects a final version to be available by the end of the month.

The model is optimized for specialized domains including cyberdefense, coding, finance, manufacturing, and electrical engineering. While Mistral claims ML4 is the most capable open-weight model developed outside of China, International reported that the system still lags behind the frontier in coding. Pierre Stock, Mistral’s VP of Science, told TechCrunch that ML4 was trained on 4,000 Nvidia GPUs, which is significantly less compute than that used by Chinese and closed-source rivals.

Mistral chief scientist Guillaume Lample stated that the model's cyberdefense capabilities would allow users to defend against threat actors who jailbreak closed models to perform attacks. International reported that the model was trained over two months on Nvidia Grace Blackwell GPUs located in European data centers. The company intends to further improve performance as it scales training capacity following its recent Series D funding round.

Outlets provided varying details on Mistral's recent financing. International and TechCrunch reported a Series D round in September valued at 21 billion euros, while Ars Technica cited a $24 billion valuation. Funding figures also differed, with Ars Technica reporting a $3.3 billion round and International reporting 3 billion euros ($3.4 billion). Ars Technica noted that Mistral’s earnings have reportedly increased 20-fold in the last year or so.