Introducing Timestral

TL;DR

Forecasting sensor readings, demand and other time-series data often requires training separate models for each dataset. Mistral’s Timestral uses mostly synthetic pre-training and jointly models related signals and known future inputs to produce forecasts with uncertainty estimates on unseen data. Mistral reports that its largest, 209M-parameter model leads three public benchmarks while outperforming larger models; all four sizes run on CPUs and support domain-specific fine-tuning.