Mistral Large 4 reaches general release after extra RL, backed by €4B European data-center bet

Mistral Large 4 (ML4) went from public preview in early October to released weights three weeks later. Today's general release follows the post-preview RL pass. Mistral has published a 30-page technical report. ML4 scores 19.7% on Cognition's FrontierCode 2.0 and 41% on Terminal-Bench 4.0, which Mistral says beats every other open-weight model with a reported score.
The training approach for cyber defence is notable. Security vendors such as CrowdStrike trained their models on years of telemetry. Mistral relied on reinforcement learning instead. "You simulate, for instance, a system that the model needs to patch, and then you reward the model whenever it manages to patch it," VP of science Pierre Stock told Computer Weekly. Mistral's pitch is that regulated industries can run ML4 on-premises to find and patch flaws without sending code to a third-party service.
ML4 does not lead everywhere. GLM-5.3 edges it on the Artificial Analysis Coding Agent Index (48.9% vs 48.3%). DeepSeek V4.1-Flash beats it on the Cyber Index (56.8% vs 50.9%).
The €4B infrastructure commitment, which follows a record €3B funding round, is Mistral's bet that Europe can build a full sovereign AI stack. Two things to watch: whether enterprise buyers actually choose self-hosted open weights over US APIs, and whether the €1B revenue target holds up.