Mistral releases Timestral, a family of time-series foundation models from 5M to 209M parameters
Mistral announced Timestral on Oct 15: four time-series foundation models (5M, 28M, 123M, 209M) for probabilistic, multivariate forecasts with past-only and known-future covariates. Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench among comparable zero-shot entries, and that a one-day forecast from three weeks of 15-minute history takes about 140 ms on an 8-thread CPU. Weights and a GitHub repo are public under a modified MIT license; fine-tuning is offered through Mistral Forge.
How you’d use it: Run Timestral zero-shot on CPU sensor or demand forecasts from GitHub.