Details
- Mistral launched a public preview of Mistral Large 4 on October 6, 2026, with API access through Mistral Studio; open weights are planned for release by the end of October.
- The natively multimodal model has approximately 1 trillion total parameters and 49 billion active parameters, making it Mistral’s largest model to date.
- Mistral says Large 4 targets coding, agentic workflows, cybersecurity, finance, law, document analysis, visual grounding, and scientific tasks. It supports more than 160 languages, including every official EU language.
- The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, which also serve the preview. The company plans deployments across multiple regions, including an independently operated European environment under European law.
- According to Mistral, Large 4 ranks among the five strongest models on the Artificial Analysis Cyber Index, scores 82% on one test, and solves 93% of Cybench challenges; the company is conducting red-team evaluations with cybersecurity partners, enterprises, and state authorities before releasing weights.
Impact
Large 4 reinforces the market’s shift toward open-weight, multimodal models that enterprises can self-host and govern, particularly for cybersecurity and regulated workloads where provider refusals or service dependence create operational constraints. If the planned October weight release delivers the reported performance, it could increase demand for sovereign AI infrastructure and steer model-development funding toward specialized, auditable systems over the next 12–24 months.
