Mistral introduced Timestral, a family of four time‑series foundation models with 5 M, 28 M, 123 M and 209 M parameters. The 209 M‑parameter version achieves top‑rank performance on the public GIFT‑Eval, TIME and fev‑bench benchmarks while using only about two‑thirds the parameters of the prior state‑of‑the‑art models. Timestral generates probabilistic multivariate forecasts (21 quantiles via a pinball loss approximating CRPS) and accepts both past‑only and known‑future covariates, treating categorical inputs natively. Its architecture combines decoupled horizon forecasting, covariate‑aware attention, synthetic‑first pre‑training (univariate processes combined into multivariate series with nonlinear dependencies) and a two‑stage training regime: initial pre‑training on purely synthetic data with an 8192‑step context, followed by post‑training on a mix of real and synthetic data with a 4× longer context. Inference runs on CPUs; forecasting one day ahead from three weeks of 15‑minute history takes 10 ms with the 5 M model and 140 ms with the 209 M model, while the 28 M variant is five times faster than the largest and the 123 M model reaches within 0.2 % of the 209 M score at three‑quarters of its latency. The models sit on the Pareto front of accuracy versus size and can be further adapted via Mistral Forge for fine‑tuning, architectural extensions, or domain‑specific data, supporting deployment on‑prem, in VPCs or air‑gapped environments under a modified MIT license.