Mistral previews Large 4, a 1T open weights model

Mistral opened a public preview of Mistral Large 4, which it also calls ML4 and, more informally, le Chonk. It is a 1 trillion parameter natively multimodal model with 49 billion active parameters, the largest and most capable model the company has built. The preview API is live today in Mistral Studio, and the weights are due by the end of the month.

The claim is not that ML4 beats every model available. Mistral says it is competitive with the strongest open source models globally and significantly outperforms any open weights model developed in the US or Europe. On what it calls critical enterprise workloads, cybersecurity, finance and law, Mistral reports it is state of the art among open models, and on visual grounding it says it goes further still, surpassing even frontier closed models. The benchmarks named on the page are DeepSWE, Terminal Bench 4.0, a cyber measure, agentic behaviour, Finance Agent and Harvey’s Legal Agent.

Training happened from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe, and the public preview is served from that same infrastructure. Mistral frames this as the point rather than a footnote: state of the art performance in critical verticals delivered through open weights, so customers control where the model runs. It will be available across multiple regions, including a European deployment Mistral operates end to end under European law, independent of other digital service providers.

The gap between preview and full release is deliberate. Until the weights land, Mistral is red teaming the model in real world settings with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities. The company’s stated reasoning is specific to security work: provider level refusals can block legitimate vulnerability research and incident response, and losing access to a capability mid incident can itself be a critical security risk.

(Original source: Mistral AI)