Mistral has released Timestral, a family of four foundation models built to forecast sensor readings, demand, financial flows and other time-series data. The models range from 5 million to 209 million parameters and can run inference on standard server CPUs rather than requiring a GPU.
Unlike a conventional forecasting model fitted to one dataset, Timestral is pretrained to work zero-shot across unfamiliar domains. It can model several related variables together, return a range of probable outcomes and use known future information such as production schedules, planned shutdowns or weather forecasts.
Mistral says the largest checkpoint ranked first across five metrics on the public GIFT-Eval, TIME and fev-bench leaderboards. These are company-reported benchmark results, and production data can behave differently. The company notes that shifting markets, aging equipment and unusual domain structures may still require adaptation.
All four models use the same architecture. On an eight-thread CPU, Mistral measured one-day forecasting from three weeks of 15-minute data at 10 milliseconds for the smallest model and 140 milliseconds for the largest. Organizations can also fine-tune Timestral on proprietary data through Mistral Forge.