Timestral: Mistral's time-series foundation models | Mistral

This post introduces Timestral, Mistral AI's first family of time-series foundation models. It details the model's architecture, training methodology (DHF, covariate-aware attention, synthetic pre-training, two-stage training with pinball loss), and performance on key benchmarks. It highlights the models' ability to perform probabilistic, multivariate forecasts with past and future covariates, and their efficiency on CPUs. The post also outlines an industrial use case and the adaptability of Timestral through Mistral Forge.