Mistral announces Timestral: SOTA time-series foundation models (mistral.ai)

Mistral has unveiled Timestral, its inaugural family of time-series foundation models designed to deliver state-of-the-art probabilistic and multivariate forecasts with only 209M parameters. This model excels in various benchmarking metrics, outperforming larger competitors while efficiently supporting past and future covariates. With sizes ranging from 5M to 209M parameters, Timestral achieves a remarkable balance of accuracy and efficiency, boasting the ability to run on standard CPUs and excel in zero-shot forecasting for diverse applications, from industrial sensors to finance. The significance of Timestral lies in its transformative approach to time-series forecasting. Traditional models like ARIMA often struggle with multiple series, while Timestral learns from a vast, diverse dataset and applies its insights across different domains, enabling superior cross-domain performance. The model employs innovative techniques such as covariate-aware attention, decoupled horizon forecasting, and a two-stage training process that optimizes learning from both synthetic and real data. This adaptability makes Timestral a powerful tool for businesses, allowing for seamless integration into existing systems and tailored fine-tuning to meet specific domain requirements. With its open-source release on GitHub, Mistral aims to enhance the forecasting capabilities of organizations across various sectors.