Mistral Opens Preview of One-Trillion-Parameter Large 4
Mistral opened a public preview of its one-trillion-parameter multimodal model, pairing frontier claims with a promise of open weights later this month.
Mistral has opened a public preview of Mistral Large 4, a one-trillion-parameter mixture-of-experts model with 49 billion active parameters. The multimodal system is available through Mistral Studio, while the company says model weights will be released by the end of October rather than on launch day.
Mistral positions Large 4 as a general model for coding, tool-using agents, multimodal understanding and professional work. It reports 61.7% on DeepSWE 1.1, 28.3% on Terminal-Bench 4, and a 59.9% score on AutomationBench, which covers workflows across applications such as Gmail, Sheets, Slack and Salesforce. The company also highlights a 93% result on Cybench and an 82% score on a vulnerability reproduction-and-patching test.
The cybersecurity positioning is central. Mistral argues that open deployment can give defenders access to powerful security capabilities without the provider refusals that block some closed models. It says Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in European data centers and supports more than 160 languages, including all official EU languages.
The claims remain largely company-reported, and the most consequential part of the announcement—the weights—has not yet arrived. Until then, developers receive API access and Mistral’s evaluation narrative, but not the ability to inspect or self-host the model.
Why it matters: Europe now has a credible flagship argument built around a combination of scale, multilingual reach, sovereign infrastructure and promised open weights. The October weight release will determine whether that argument becomes a durable ecosystem advantage or remains an impressive preview.