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

Mistral AI released a public preview of Mistral Large 4, a new flagship model the company says is the strongest open-weight AI system built outside China. The model, which Mistral calls ML4, has 1 trillion total parameters and 49 billion active parameters, the smaller subset the model actually draws on when answering a given query. It is natively multimodal, meaning it was built from the start to handle text and images together rather than having vision added on afterward. Developers can test it now through Mistral’s API on a platform called Mistral Studio, though Mistral says the full model weights, the files needed to run the model independently, will not be released until the end of October 2026.

Mistral says ML4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs inside the company’s own data centers in Europe, and the preview runs on that same infrastructure. According to Mistral, the training data spanned more than 160 languages, including every official language of the European Union.

Cybersecurity results

Mistral says ML4 ranks among the top five models worldwide on the Artificial Analysis Cyber Index, an independent evaluation of how well AI models find and fix security flaws in real software, and leads open-weight models built outside China by a wide margin. On one test in that index, which asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scored 82%, the highest score Mistral reports for any model on that test. It also solved 93% of the 40 challenges in Cybench, a benchmark drawn from security competitions.

Mistral says that score matters because several closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same vulnerability test. According to Mistral, that isn’t because those models lack the ability; it’s because their safety filters refuse the task. Mistral argues that defenders often need to prove a flaw is real before they can fix it, and that refusals built into closed models can get in the way during an active security incident.

Beyond cybersecurity, Mistral reports strong results on coding and office-style tasks. The company says ML4 scored 61.7% on a coding benchmark called DeepSWE v1.1 and 59.9% on AutomationBench, a test covering 657 business workflows across apps such as Gmail, Google Sheets, Slack and Salesforce, putting it ahead of China’s DeepSeek V4 Pro on both measures, according to Mistral’s own benchmarking. In a blind evaluation Mistral commissioned from human annotators at Surge AI, ML4 ranked second out of five models on coding quality, behind only Anthropic’s Claude Opus 5.

Mistral has not yet published the model’s architecture or post-training methodology, saying more details will follow as it finishes preparing the weights for release. Until those open weights ship at the end of October, outside researchers have no way to independently verify Mistral’s benchmark claims, including the cybersecurity scores the company is promoting most heavily.