Mistral AI has released Timestral, its first family of time-series foundation models, for probabilistic, multivariate forecasting, the company said on 15 July 2026.
The models come in four sizes, from 5 million to 209 million parameters, and run efficiently on CPUs. Mistral said the family sits on the Pareto front of accuracy versus model size, and the largest ranks first on GIFT-Eval, TIME and fev-bench, beating the previous state of the art with two-thirds of the parameters.
Timestral supports past-only and known-future covariates and zero-shot forecasting. It combines synthetic pre-training, covariate-aware attention and long-context learning.
Organisations can fine-tune it for their own domains through Mistral Forge, the training pipeline Mistral uses for its language models.
The announcement was published on 15 July 2026. Mistral did not give pricing for Forge fine-tuning.