Mistral releases Timestral forecasting models
Four time-series models, from 5M to 209M parameters, for probabilistic forecasts that can run on a CPU.
Official source (opens in a new tab)
Mistral released Timestral on 15 October 2026, its first family of time-series foundation models. The four sizes are 5 million, 28 million, 123 million, and 209 million parameters. Mistral says the 209 million parameter model ranks first among pretrained, zero-shot models on GIFT-Eval, TIME, and fev-bench, with about two-thirds the parameters of the previous best. Forecasts are probabilistic and multivariate, with past-only and known-future covariates, a context of 32,768 time steps, and 21 quantiles. The licence is a modified MIT, and Mistral says the models can run on-prem, in a VPC, or air-gapped, including on a CPU.
Mistral says a GitHub repository is up so you can run Timestral on your own data, and that domain adaptation is available through Mistral Forge. A technical report is still to come.
Who it's for: people forecasting sensors, demand, or other series, not people writing or coding in chat.
Fits: none of the current task lists. It is a forecasting model, not a writing, research, or coding tool.