Mistral Releases Large 4 Weights After Three More Weeks of Reinforcement Learning

French AI developer Mistral opened downloads of Mistral Large 4’s weights Thursday, Oct. 29, after another three weeks of reinforcement learning, making its natively multimodal model with one trillion parameters available to developers.

Mistral is positioning the model for coding agents and defensive cybersecurity. Its Oct. 6 preview had set expectations for a weight release later in the month.

Large 4 scored 48% on the Coding Agent Index and used 47% fewer tokens per task on average than GLM-5.3. It scored 41% on Terminal-Bench 4.0, 13 percentage points above its preview result. The Coding Agent Index combines three benchmarks and evaluates agent configurations, so its score reflects the software directing a model as well as the model itself.

Large 4 scored 51% on the Cyber Index, which combines evaluations of vulnerability discovery, reproduction and validation, and patching.

The weights are available under Mistral’s Ecosystem Open-Weight License. The license restricts some model-as-a-service offerings.