Timestral

Timestral is an open-source family of time-series foundation models developed by Mistral AI for multivariate probabilistic forecasting. It predicts multiple related variables simultaneously while incorporating historical and known-future information. Four model sizes are available, ranging from 5 million to 209 million parameters. Its causal Transformer architecture uses Decoupled Horizon Forecasting to generate predictions efficiently. Each forecast produces 21 quantiles to represent uncertainty rather than a single expected value. The models support numerical and categorical covariates, missing observations, and context windows of up to 32,768 time steps. Standalone Python inference runs on CPUs or GPUs using PyTorch. Timestral is designed for forecasting research and practical predictive analytics.