A promise is not an artifact.
Mistral launched Mistral Large 4 as a public preview on October 6. Its launch post says the model has 1 trillion total parameters and 49 billion active parameters. Its model documentation says 1.05 trillion total parameters, 52 billion active parameters, and a 1.6 billion parameter vision encoder. The preview API is live under mistral-large-4. Mistral says the weights will be released by the end of the month.[1][2]
That is an interesting release. It is not yet a weight drop. The difference matters if your plan depends on private deployment, a fixed checkpoint, a license review, quantization, or a hardware fit test. Today you can test the provider route. You cannot inspect or run the promised checkpoint because Mistral has not published it.
Open-weight describes custody after delivery. It does not make a preview API self-hosted.
The spec sheet already has drift.
Two first-party pages disagree on the active parameter count. The launch post says 49 billion. The model page says 52 billion. They also describe the total count at different precision. We cannot resolve that conflict from the published pages.
This does not prove that the model route changed. It proves that a copied parameter badge is weak intake evidence. Record the source URL and read date beside each number. Leave the conflict open until Mistral aligns the pages or the checkpoint metadata settles it.
Public preview can move under your test.
Mistral's lifecycle policy says public-preview models may receive silent updates. It also says a preview has no guaranteed path to general availability and may be retired with one month's notice. General-availability models are the stage where Mistral promises no silent updates.[3]
Pinning mistral-large-4 records the route name, not immutable behavior. Save the date, request settings, tool schema, output, token use, and cost beside every preview result. If the same replay changes later, you need enough evidence to tell model drift from harness drift.
The current numbers describe the hosted route.
Mistral lists a one-million-token context window and preview prices of $1.36 per million input tokens and $4.18 per million output tokens before the temporary discount shown on its model page. The company reports a 49.8 percent Coding Agent Index score built from three coding and terminal evaluations. Mistral notes that those results were evaluated before the harness became public.[1][2]
Artificial Analysis currently classifies the preview as proprietary because the weights are unavailable. Its own page reports a 38 Intelligence Index score, 116.1 output tokens per second, and much higher token use than the median of its comparison class. It lists a 524k context window, which does not match Mistral's one-million-token documentation.[4] Treat every limit as a route-specific fact to test, not a number to copy into procurement notes.
Run two acceptance lanes.
The API lane can start now. Replay work your team knows, record the exact harness and settings, inspect patches, and compare total cost and time. Include a long-context probe if that limit matters to your workload.
The custody lane starts when a real checkpoint, license, model card, and download location exist. Then verify the files, calculate the memory budget, choose a quantization, exercise the target runtime, and rerun the same tasks. A hosted preview result does not predict local throughput or deployment cost.