TimesFM (Time Series Foundation Model) (github.com)

Google Research released TimesFM 2.5, a pretrained, decoder-only time-series foundation model for forecasting (ICML 2024). The open-source checkpoints are on Hugging Face and the model is also available in BigQuery; the repo and package are community-maintained (not an “official” Google product). Compared to TimesFM 2.0, 2.5 reduces model size to 200M parameters (from 500M), boosts context length dramatically to 16k tokens (from 2,048), and supports optional continuous quantile forecasting up to a 1k horizon via a 30M-parameter quantile head. The release removes the frequency indicator, adds new forecasting flags, and ships an upgraded inference API with a forthcoming Flax build for faster inference and restored covariate support. This matters because TimesFM applies the foundation-model paradigm to structured time-series tasks: long-range context, compact parameterization, and built-in probabilistic outputs enable transfer across domains and scalable deployment. The code supports PyTorch and Flax, runs on CPU/GPU/TPU/Apple Silicon, and exposes a ForecastConfig for normalization, flip invariance, positivity inference, and quantile-crossing fixes. Example usage loads the pretrained model from Hugging Face and returns point and quantile forecasts (shape: batch × horizon × quantiles). TimesFM 2.5 therefore offers a practical, efficient baseline for both research and applied time-series forecasting at scale.