Mechanism: The model is available through Mistral Studio’s public API preview, while weights are scheduled for a later release. That separates immediate API testing from self-hosting: developers can evaluate hosted access now, but cannot infer from this announcement that they can inspect or deploy the weights themselves.
Observed facts: TestingCatalog describes Large 4 as a natively multimodal mixture-of-experts model with 1 trillion parameters and 49 billion active parameters. The article gives preview API prices of $1.36 per million input tokens and $4.18 per million output tokens. It also reports 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench, attributing the performance claims to the announcement and evaluations it cites.
Why now: Mistral has moved from a reported upcoming release to public preview access, while the later weight release remains a separate milestone.
Implications: Readers can test API behavior and budget against stated token prices now. Claims about benchmark standing, customer control, or self-deployment should wait for comparable independent testing and actual weight availability.
Evidence boundary: The supplied public source is TestingCatalog’s account of the launch and reported results; this record does not independently verify benchmark protocols, scores, or the future weight-release date.
Watch: Whether Mistral releases the weights by the stated end-of-October target, and whether independent evaluations reproduce the reported results.