Mistral Large 4 is available in public API preview, but its weights are not yet downloadable. Announced on October 6, 2026, the multimodal mixture-of-experts model combines language, vision, coding, and agent capabilities in one system. Mistral says it plans to release the weights by the end of October; until then, teams can try the hosted preview through Mistral Studio.
That distinction matters: this is a substantial model announcement, but it is not yet a self-hosting release. Teams evaluating open models can test the API now while postponing infrastructure and deployment decisions until the weights and fuller technical details arrive.
Mistral Large 4 at a glance
| Detail | What Mistral currently says |
|---|---|
| Availability | Public API preview, announced October 6, 2026 |
| Model name in API docs | mistral-large-4-0 |
| Architecture | Multimodal mixture of experts |
| Parameters | 1.05 trillion total; 52 billion active in the model documentation |
| Context window | 1 million tokens in the model documentation |
| Weights | Mistral says it expects to release them by the end of October |
Mistral’s launch post reports 49 billion active parameters, while the model documentation lists 52 billion. That discrepancy is worth noting; verify the current documentation rather than treating either number as settled for capacity planning.
What is notable about this release?
Mistral is positioning Large 4 for coding, multi-step agent work, image understanding, and professional tasks such as working with spreadsheets and documents. Its announcement reports results across software-engineering, cybersecurity, business-workflow, legal, finance, and multimodal evaluations.
These are useful directions for an evaluation plan, not proof that the model will outperform alternatives on a team’s own workload. The scores are largely reported by Mistral, some evaluations are internal or narrowly scoped, and the public preview is still being refined. The company says it will share more about the architecture, benchmarks, and post-training as it works toward the weight release.
What the preview means for developers
The practical next step is a bounded API evaluation rather than a production migration. Choose tasks that reflect your actual product: for example, repository changes, document extraction, image-grounded questions, or tool-using workflows. Compare successful completion, corrections, latency, token use, and cost against the model you already run.
For agent use, keep the same operational controls you would use with any model: narrow tool access, validate outputs against application rules, log actions, and require human approval before consequential external or production changes. A high score on an agent benchmark is not a substitute for permissions and task-level tests.
If self-hosting, data locality, or model customization is the deciding factor, wait for the weights and accompanying license, hardware guidance, and deployment documentation. “Open-weight” describes the announced direction; it does not mean those artifacts are already available today.
What to watch next
Mistral says it plans to publish the weights by the end of October 2026, along with further architectural and training details. When they arrive, teams should review the actual license and model artifacts, check supported inference stacks, measure memory and throughput needs, and repeat their own task and safety evaluations. The preview results may change as the model is trained further.
For help comparing models or integrating a preview safely, see AI consulting and custom AI development. We also help build AI-enabled websites and applications around measured, reviewable workflows.
Official sources
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4 is Mistral AI's new general-purpose multimodal model, built with a mixture-of-experts architecture. Mistral's model documentation lists 1.05 trillion total parameters and 52 billion active parameters.
Can I use Mistral Large 4 now?
Mistral announced a public preview API on October 6, 2026, through Mistral Studio. Its model documentation lists the API model name as mistral-large-4-0. Access and terms can change during preview.
Can I download and run Mistral Large 4's weights now?
Not at launch. Mistral said it expects to release the model weights by the end of October 2026. Until then, the public preview is available through its hosted API.
How should teams interpret Mistral Large 4 benchmark results?
Treat them as an initial screening signal, not a guarantee. Many results in the launch announcement are Mistral-reported evaluations, and the preview is still being refined. Test representative tasks, safety, latency, and total cost in your own environment.