Timestral

Mistral’s Timestral is a family of four probabilistic time-series forecasting models, sized 5.2M, 27.8M, 122.6M, and 209.3M parameters. Its central innovation, Decoupled Horizon Forecasting, separates prediction tokens from causally encoded context: the model can train forecasts at multiple distances while keeping observed values available to inform later predictions. A role-aware attention framework also lets forecasts use known-future covariates without exposing future target values. Timestral is pretrained exclusively on synthetic data, then post-trained on real and synthetic data with longer context. Across GIFT-Eval, TIME, and fev-bench, Timestral-209M ranks first on all six reported metrics, with Timestral-123M second; the family’s results place every size on the accuracy-versus-parameter Pareto front. For example, Timestral-209M records a GIFT-Eval MASE of 0.6575 and CRPS of 0.4460, and a TIME MASE of 0.6262 and CRPS of 0.5223. On fev-bench, its SQL skill score is 50.39%. The paper compares Timestral with forecasting models including Chronos-2 and Toto-2.0, and reports that forecasts up to 1,024 time steps can be produced in one forward pass.