Mistral introduced Large 4 in API public preview on October 6, 2026. The model is available for evaluation through Mistral’s service, while a weight release is planned for later in the month. As of this announcement, it should not be described as a downloadable open-weight release. Mistral’s launch post supplies the release status.
Public preview and open weights are separate milestones
The model documentation lists the API model as mistral-large-4 and labels it public preview. Mistral’s launch discusses continued safety work before the planned distribution of weights. That sequence means an API evaluation can begin before teams can assess self-hosting requirements.
Do not build a deployment plan around an expected download date alone. Self-hosting would require the actual license, checkpoint, inference support and hardware assessment. Until those are available, a team can evaluate the hosted model while keeping the later operating model as an unresolved decision.
Evaluate the task, rather than the headline architecture
Large-model announcements often foreground parameter counts and benchmark results. Those can describe the vendor’s design and evaluation, but they do not predict success on a company’s own documents or tool workflows. A public-preview model also deserves version tracking so results can be interpreted when behavior changes.
Use examples that include the intended output structure, tools and failure conditions. For a coding task, inspect whether the change preserves existing behavior and passes relevant checks. For document analysis, verify that cited evidence supports the answer. Record cost and time alongside acceptance; a superficially cheaper request can become expensive if it needs repeated repair.
Keep the preview reversible
A bounded trial should avoid making the model the sole dependency for a critical workflow. Keep evaluation inputs and outcomes available for comparison with the current model, and make refusal or malformed output visible. The application should own authorization and validation rather than expecting the model to enforce those boundaries.
The October 6 release is significant as a new frontier-model candidate from Mistral. The immediate decision is whether its API preview improves a measured workflow. A separate decision about open weights should wait for the actual release materials, not assume that preview access already provides deployment freedom.