Mistral has released its first time-series foundation-model family. Timestral produces probabilistic multivariate forecasts using historical data plus past-only or known-future covariates such as production schedules.
Four sizes range from 5 million to 209 million parameters, with 32,768 time steps of context and 21 output quantiles. Code and models are available, and inference can run on CPUs. Mistral offers adaptation through Forge, including on customer infrastructure.
Mistral reports leading results on GIFT-Eval, TIME and fev-bench; those rankings do not guarantee performance on changing industrial data. Licensing has two layers: the inference code is Apache 2.0, while the model weights use Modified MIT, which does not authorize use when the company’s or employer’s global consolidated revenue exceeded $20 million in the preceding month; a separate commercial license must then be requested. The full technical report is still forthcoming.
Foundation models extend beyond text. Numerical forecasting becomes a component agents can call rather than imitate with prose.