Timestral: Mistral release time-series foundation models

Timestral is a family of time-series foundation models that produces probabilistic, multivariate forecasts while remaining compact and CPU-efficient. Offered in four sizes (5M, 28M, 123M, 209M parameters) with a 32,768-step context and 21 quantile outputs, it ranks first on the GIFT-Eval, TIME and fev-bench leaderboards for pretrained zero-shot models and sits on the Pareto front of accuracy versus model size, outperforming larger baselines such as TimesFM-3 and Falcon-2.0. Timestral supports multivariate targets, past-only and known-future covariates (including categorical inputs), and returns full quantile sets for risk-aware decision-making. Inference is CPU-friendly: one-day-ahead forecasts from three weeks of 15‑minute data take about 10 ms with the 5M model and 140 ms with the 209M model on an 8-thread server, and each model size trades accuracy for latency to suit deployment needs.

The design combines several technical advances: decoupled horizon forecasting (keeping full history in-sequence while using learned forecast tokens), covariate-aware attention that prevents target-future leakage, and a synthetic-first, two-stage training pipeline (pretrain on synthetic data with context 8,192 then post-train on mixed real/synthetic data with 4× longer context) using pinball loss across 21 quantiles. Timestral is intended for production adaptation via fine-tuning, architectural extensions and controlled data mixes, and can be deployed on-premises, in VPCs or air-gapped environments and integrated into monitoring and agentic workflows for anomaly detection, root-cause analysis and automated response.