Mistral previews a trillion-parameter AI model trained on European infrastructure
Mistral has opened an API preview of Large 4, a multimodal model with 1 trillion parameters and 49 billion active per token; it says the model was trained on 3,800 NVIDIA GPUs in its European datacentres. The French company says the model performs strongly in coding, business tasks and cybersecurity, and plans to release its weights by the end of October, giving organisations the option to run it on their own infrastructure.
Bottom line — Benchmarks of Mistral’s latest release show it scores lower than foreign rivals, but advances the frontier of European AI.
Go deeper 5
-
Mistral says Large 4 ranks among the top five models globally on the Artificial Analysis Cyber Index and scored 82% on a test reproducing and patching a software vulnerability.
-
The company reports a 93% score on Cybench and says the model refuses malicious cybersecurity requests more often than other open-weight models it assessed.
-
Mistral says Large 4 scored 59.9% on AutomationBench, which tests 657 business workflows, and 1,393 Elo on AA-Briefcase, a benchmark for longer knowledge-work tasks.
-
The API preview is available now, but the weights are not. Explainx.ai notes that the published benchmarks are vendor-reported and need independent evaluation.
-
Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs and that its training data spans more than 160 languages, including every official EU language.