Mistral previews Large 4, its trillion-parameter multimodal model

Mistral AI has opened a public preview of Large 4 on Mistral Studio, with open model weights scheduled for release at the end of October, AI Magazine reports. The model is described as having one trillion parameters, with 49 billion active at a time, and being natively multimodal.

Mistral AI Watch analysis

What happened

AI Magazine says the preview is available through Mistral Studio. The model’s full open-weight release is scheduled for later this month, so the preview and the downloadable weights are distinct milestones. Mistral’s Applied AI Engineer Lead Nicolas Schuhl describes Large 4 as the company’s strongest model for tasks including coding, agentic workflows and multimodal understanding; that is Mistral’s assessment, not an independent benchmark in the report.

Why it matters

Large 4 puts a clear scale and capability claim behind Mistral’s next flagship: one trillion total parameters, but 49 billion active parameters. For developers and organisations, the near-term question is what the Studio preview can do in practice; the open-weight release could later allow more direct access to the model’s weights and customisation.

That timetable makes this more than a model teaser, but the evidence here does not establish independent performance results or the terms of access to the preview. The end-of-month weights are the milestone to watch, not a feature already in hand.

Our read

A public preview now, weights later: useful news with a meaningful next step attached. The size figures are striking, but they are not a substitute for task-level results. We’d judge Large 4 on how it performs when people can try it, and on whether the promised open weights arrive as scheduled.

What to watch

• Whether Mistral releases the open weights by the end of October.
• What access the Mistral Studio preview offers and which capabilities users can test.
• Independent results on coding, agentic workflows and multimodal tasks.

Discussion spark: When the weights arrive, what should matter more in judging Large 4: its performance on practical tasks, or how freely developers can customise and run it?

not affiliated with or endorsed by Mistral AI