Mistral AI Releases Timestral Time Series Foundation Models
Mistral AI has launched Timestral, a family of patch-based transformer foundation models designed for zero-shot time series forecasting, available in parameter sizes ranging from 5M to 209M. The models support both univariate and multivariate forecasting, long contexts of up to 32,768 historical steps, and probabilistic predictions across 21 quantiles. This release expands Mistral AI's portfolio beyond text and language models into specialized numerical forecasting, bringing advanced transformer-based capabilities to enterprise time series analytics. Open-weight availability allows researchers and developers to easily run zero-shot forecasts and reproduce benchmarks across diverse industrial domains. The flagship 209M parameter model features a 24-layer patch-based transformer architecture with an embedding dimension of 768, processing 32 time steps per patch. It can forecast up to 1,024 steps in a single forward pass, while handling longer horizons block by block using PyTorch on either CPU or GPU.
## BACKGROUND
Time series forecasting involves predicting future values based on previously recorded data points, which is crucial for domains like finance, energy, and supply chain management. Zero-shot forecasting allows a pre-trained model to make accurate predictions on entirely new datasets without requiring task-specific fine-tuning. Covariates are external secondary variables—such as weather conditions or promotional calendars—that can influence the primary target values being forecast.