Mistral previews Mistral Large 4, a 1-trillion-parameter open-weight model

Mistral released Mistral Large 4 (ML4) as a public preview on October 6. The multimodal model has 1 trillion parameters but uses only 52 billion of them for each token, and the company calls it the largest and most capable model it has built. The preview API is available in Mistral Studio, and the weights, which anyone will be able to download and run, are due at the end of the month.

Mistral says the model excels at coding, agentic work and image understanding, and leads open models on enterprise work such as security, finance and law. On a security test that asks a model to reproduce and then patch a real vulnerability in open-source software, it scored 82%, the highest of any model, while some closed models scored near zero because they refused the task.

ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, on data covering more than 160 languages, including every official EU language. A European deployment that Mistral runs end to end under European law, independently of other providers, targets companies and institutions that want control over their AI.

The performance figures come from Mistral, with only some from outside evaluators. Until the weights are out, Mistral is red-teaming the model in real-world settings with security firms, vetted partners and government agencies, and it plans to publish architecture details and more benchmarks alongside the weights.