On October 6, Mistral made Large 4 available as a public API preview. The model handles text and visual material; the company plans to release weights by the end of October. We treat the results in its announcement as developer claims, not universally established superiority.

Why might this matter beyond leaderboards? Imagine a technician comparing a drawing, an order and a service record. A useful assistant would identify the supporting material, not merely propose an answer, and leave the decision to a person. Such a workflow could reduce copying and searching if it passed tests on the organization's own material.

The prospect of self-hosting has a different value from a high score: an organization could gain control over where data goes and when its system changes. That requires released weights, clear licensing, sufficient hardware and an operations team. A remote-service preview does not fulfil those conditions by itself.

An initial trial should be small and reversible. Select anonymized documents, prepare correct answers in advance and track missing details, false conclusions and total task cost. Do not authorize the agent to approve contracts or change production independently. A fluent answer is not a correctness check.

Our optimistic editorial estimate is 2–6 weeks for a limited API pilot if an integration team is ready. This is not Mistral's timetable. We will not estimate a safe self-hosting date yet: it depends on the weights, their terms and the actual resource requirements.