Presentiamo Timestral

Mistral has released Timestral, its first family of foundation models designed for time-series forecasting. The largest model, with only 209 million parameters, achieves state-of-the-art results on key industry benchmarks, outperforming much larger models. Timestral supports multivariate and probabilistic forecasting, can incorporate covariates, and is capable of zero-shot predictions on unseen data. The models are highly efficient, able to run on standard CPUs, and can be fine-tuned or integrated into complex, multi-modal workflows for applications like industrial monitoring.