Mistral pitches open-weight Large 4 for on-prem cyber defence, backed by 30-page technical report

Weeks after releasing the weights, Mistral has a sharper sales pitch for Large 4: cyber defence and sovereign AI. Mistral wants security teams and regulated industries to run the model on their own infrastructure and use it to find and patch flaws in their code without sending data to a third-party service, Computer Weekly reported on November 13. ML4 is a multimodal mixture-of-experts model with one trillion total parameters, about 52 billion active at a time. It reached public preview in early October and its open weights were released on October 29, followed by a 30-page technical report.

The report gives the numbers. ML4 scored 19.7% on Cognition's FrontierCode 2.0 and 41% on Terminal-Bench 4.0, beating every other open-weight model with a reported score. Three weeks of extra reinforcement learning after the preview raised Terminal-Bench by 13 percentage points and cut output tokens by 30%. Mistral VP of science Pierre Stock explained the method: 'You simulate, for instance, a system that the model needs to patch, and then you reward the model whenever it manages to patch it, and you penalise the model whenever it's stuck in a dead end.' Security vendors such as CrowdStrike train on years of telemetry; Mistral relied on RL instead.

ML4 is not the best open model everywhere. GLM-5.3 edged it on the Artificial Analysis Coding Agent Index, 48.9% to 48.3%. DeepSeek V4.1-Flash led more clearly on the Artificial Analysis Cyber Index, 56.8% to 50.9%. The model costs about $3.48 per million tokens through Mistral, and Mistral Studio is meant to simplify deployment across cloud, VPC and on-prem setups.

The sovereignty argument is Mistral's real advantage as a European alternative to US closed models, especially as Meta appears to retreat from open weights with the hosted-only Muse Ember. Caveats: frontier closed models still lead on most coding benchmarks, and running a 1T-parameter model on-prem takes serious GPU capacity. NVIDIA had to split GLM-5.3 across two DGX Stations just to serve it locally.