MistralAI Releases Timestral-209M for Advanced Time Series Forecasting

MistralAI has released Timestral-209M, a 209 million parameter model designed for zero-shot time series forecasting. This model supports univariate and multivariate forecasting, handles past-only and known-future covariates, and provides probabilistic predictions across 21 quantiles. It boasts a context window of up to 32,768 historical time steps and can forecast up to 1,024 steps in a single forward pass. Timestral-209M is available for use via PyTorch on both CPU and GPU, and its source code can be accessed on GitHub. According to the model card, Timestral-209M achieves state-of-the-art zero-shot accuracy across major forecasting benchmarks like GIFT-Eval, TIME, and fev-bench, outperforming other models in metrics such as MASE and CRPS.