Mistral previews 1.05T-parameter Large 4 and says weights will follow this month

Mistral has opened a public preview of Mistral Large 4, its largest model to date and a new multimodal mixture-of-experts system aimed at coding, agentic workflows and enterprise reasoning. Mistral's model documentation lists 1.05 trillion total parameters, 49 billion active parameters and a 1.6-billion-parameter vision encoder. The same documentation lists a one-million-token context window, while the launch post describes the model as natively multimodal and fluent in more than 160 languages. Those specifications make the preview important for the open-weight landscape, but they do not yet constitute a downloadable weight release.

The release state matters for the registry. Mistral says the preview API can be used now in Mistral Studio and that the weights will be released by the end of October. At the time of this brief, Open Model Weights therefore treats Mistral Large 4 as an announced open-weight model rather than a field-verified weight record. No repository artifact, shard list or final weight-license evidence should be inferred from the preview language. Once the files are public, each of those fields can be checked directly and independently.

Mistral says Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. The company also publishes benchmark comparisons across coding, cybersecurity, finance, legal and multimodal tasks. Those performance numbers remain publisher-reported until independently reproduced. For the evidence layer, the durable facts today are the architecture and context specifications in Mistral's own documentation, the current API-preview status and the explicit commitment to publish weights later this month.