Introducing Timestral

Mistral has released Timestral, a family of time-series foundation models designed for probabilistic, multivariate forecasting. The largest model, at 209M parameters, reportedly outperforms larger models on key public benchmarks while remaining CPU-efficient. Available in four sizes, the models excel at zero-shot forecasting and can incorporate past, future, and categorical covariates. Key technical contributions include a synthetic-first data approach, covariate-aware attention, and a decoupled horizon forecasting method.