Science Behind the Open Cyber Frontier with Mistral Large 4
Mistral releases 1-trillion-parameter open-source Large 4 trained on its own European compute
“Ultimately, our goal is for the best model in the world to be open source.”
Mistral has released Large 4 (Lechonk), a ~1 trillion parameter open-weight model — roughly one-third larger than their previous 600B ML3 — trained on Mistral's own GPU cluster in Brie-Luchatel, France. The launch marks Mistral's first step into frontier-scale training on owned infrastructure, signaling a deliberate move to compete with closed models at the top of the capability curve. Highlighted strengths include cyber defense, finance, law, and multimodal vision, with the open-weight release intended to unlock community-driven specialization.