Mistral AI has released the weights of an updated Mistral Large 4, its 1-trillion-parameter multimodal model. The company says the update improves coding, defensive cybersecurity and reasoning efficiency. It follows a public preview released earlier this month.
Mistral says it made the gains after three weeks of additional reinforcement learning. On the Artificial Analysis Coding Agent Index, the model scored 48% while using 47% fewer tokens per task on average than GLM-5.3, which the company says had an effectively tied score. On Terminal-Bench 4.0, its score rose 13 percentage points from the preview to 41%, while its output token use fell 30%.
The company says Mistral Large 4 now supports three reasoning modes, giving developers a choice between performance and cost. It has released the model’s weights for organizations that want to customize or host it themselves.
Read nextNVIDIA releases Nemotron 3.5 Transcribe and Granary v2 on Hugging FaceMistral reports security test results
On the Artificial Analysis Cyber Index, which tests vulnerability discovery, reproduction and patching, Mistral Large 4 scored 51%, according to the company. Mistral also reports an average refusal rate of 96.6% on malicious cybersecurity requests in its tests across JailbreakBench, StrongREJECT and AgentHarm.
Mistral says it observed the model refusing black-box attacks against live targets while monitoring traffic after the preview. It also says it sees potential for the model to help conduct real attacks. The company is giving vetted enterprise design partners, including Orange Cyberdefense, preview access to a specialized variant for advanced security work. Testing of that variant is restricted to owned or explicitly authorized systems in controlled environments.
The updated Mistral Large 4 weights are available now. Mistral says it is modifying the model’s guardrails and will continue updating them in the months ahead.
