Pioneering Time Series Foundation Models with Tiny-Scale Architectures

The emergence of time series foundation models (TSFMs) has enabled accurate forecasting with minimal fine-tuning. By leveraging knowledge learned during the pre-training phase, such as patterns of trends and seasonality, TSFMs also support zero-shot forecasting, greatly expanding practical use cases. This presentation will begin with an overview of TSFMs, covering representative examples, key datasets, and common insights. It will then focus on a compact model that the presenter contributed to, highlighting its architecture and time series-specific innovations. This model family, which includes an ultra-lightweight TSFM, has been downloaded over 21 million times and is gaining substantial attention in the time series domain. Despite its small size, it outperforms several recent transformer-based foundation models. Moreover, its lightweight design allows inference and fine-tuning entirely on CPUs, delivering practical benefits for real-world applications.