Mistral opened a public API preview of Mistral Large 4 on 6 October 2026. Its release announcement says developers can try the model through Mistral Studio, while downloadable weights are planned for later in October. For public developers, this is an opportunity to test a hosted model, not a public release of weights for self-hosting.
What developers can test now
Mistral’s model documentation lists version 26.10 as a public preview under the API identifier mistral-large-4. It describes a mixture-of-experts model with 1.05 trillion total parameters and 49 billion active parameters. The active figure refers to the portion used for a token; it does not describe the size of a complete downloadable checkpoint.
The same documentation lists multimodal input, a 1.6 billion-parameter vision encoder, a one-million-token context window, structured outputs and function calling. These are Mistral’s published specifications and interface options, not independent measurements of quality or reliability across those workloads. A long advertised context window, for example, does not establish accurate recall at every position.
What is still pending
Mistral says it intends to publish the weights by the end of October. Its Hugging Face page is marked as an upcoming release rather than a downloadable checkpoint. The page currently gives an expected date of 31 October, but that is a plan, not a completed release.
Until the public files and applicable license are available, teams relying on the public release cannot verify its self-hosting terms, hardware requirements or deployment behavior. Describing the preview as open-weight without this timing distinction would obscure the difference between a hosted API test and operational control of the model.
The practical decision
Developers considering a pilot can test representative text and image inputs, tool calls and structured responses through the preview, then measure their own quality, latency, cost and safety behavior. Mistral’s performance rankings in its announcement are vendor claims and do not replace workload-specific testing. A self-hosting decision should wait for the actual checkpoint, license and deployment guidance.
