tl;dr: Mistral released Timestral, a family of time-series foundation models (5M–209M parameters) that tops the GIFT-Eval, TIME, and fev-bench leaderboards for zero-shot forecasting, beating larger models like TimesFM-3 and Chronos-2. The models deliver probabilistic multivariate forecasts, support past and known-future covariates, and run on CPU (10–140 ms per forecast). Key technical contributions include Decoupled Horizon Forecasting, covariate-aware attention, and a synthetic-first two-stage training approach, with fine-tuning available via Mistral Forge under a modified MIT license.