Mistral announced Timestral on October 15 as its first family of time-series foundation models. The family comes in 5M, 28M, 123M and 209M parameter sizes and is designed for probabilistic multivariate forecasting, including past and known-future covariates.
Small models designed to run on CPUs
Mistral says the models can run efficiently on CPUs and can be deployed on premises, in a private cloud or in air-gapped environments. It lists a 32,768-step context length and a modified MIT license, and says the larger sizes trade more compute for stronger forecast accuracy.
Mistral reports that the largest model ranks first across GIFT-Eval, TIME and fev-bench, and says the results have been submitted to those public leaderboards. The ranking is therefore attributed to the company and its submitted evaluations; it is not described here as independently reproduced.