Mistral releases Timestral time-series foundation models

Mistral released Timestral on 15 October 2026, which it calls its first family of time-series foundation models. The four sizes are 5M, 28M, 123M and 209M parameters. Mistral says the largest, at 209M parameters, ranks first on GIFT-Eval, TIME and fev-bench among pre-trained, zero-shot models, outperforming the previous state of the art with only two-thirds of the model parameters. The models produce probabilistic multivariate forecasts, with 21 quantiles, a context of 32,768 time steps, and past-only and known-future covariates. Mistral says that on a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model. The licence is Modified MIT. Fine-tuning is available through Mistral Forge.

What it means for business: a small forecasting model Mistral says can run on a CPU. The benchmark ranking is the company's claim. Read the Modified MIT licence before commercial use.