Supporting case
Timestral provides downloadable forecasting models small enough for CPU use—with a material licence limit
What happened
Mistral released Timestral on October 15; it is recovered older material, not a November 4 release. The models forecast numerical sequences, such as a plant’s sensor readings or demand over time. They are not general-purpose chat assistants. Four sizes range from about 5 million to 209 million parameters. The official repository contains downloadable weights, and the instructions support CPU or GPU inference. Model card ↗; actual files ↗.
Mistral reports fast CPU forecasts and favorable benchmarks; those performance claims are supplier evidence, not independently measured benefits in a small institution. The model’s Modified MIT licence also matters. Its granted rights are unavailable when the company’s or employer’s global consolidated revenue exceeded $20 million in the preceding month. Those users need a separate commercial licence at Mistral’s discretion or another permitted access route. The code repository’s Apache licence does not replace the weights’ terms. Model licence ↗.
For The Race, this is a concrete partial alternative to purchasing hosted forecasting inference: an eligible institution can possess and operate the component. Integration skills, forecasting reliability and licensing still determine usable independence. The current landscape did not contain this release; the existing Mistral Large 4 award concerns a separate model and access change.
Dates and implementation stage
Action/stage 2026-10-15; reporting 2026-10-15; approved 2026-11-05
