google-research/timesfm

Google's TimesFM is a 200-million-parameter foundation model that matches state-of-the-art supervised forecasting methods in zero-shot. No fine-tuning. No training on your data. You feed it a time series, it predicts the future. Pretrained on 100 billion real-world time points across retail, finance, energy, weather, and more. The architecture is decoder-only, inspired by LLMs but purpose-built for time series. Where GPT predicts the next token, TimesFM predicts the next time step. The key insight from Google Research: you can treat time series forecasting the same way you treat language modeling, and the transfer learning works across completely different domains. TimesFM 2.5, the latest version, adds covariate support through XReg (exogenous regressors), letting you include external variables like holidays, promotions, or weather in your forecasts. It also introduces fused QKV matrices for faster inference. Available on Hugging Face and natively in Google BigQuery, which means you can run forecasts directly inside your data warehouse without exporting anything. The zero-shot performance is the headline. On multiple benchmark datasets across different domains and temporal granularities, TimesFM comes close to or matches supervised approaches that were specifically trained on each dataset. A 200M parameter model doing this is remarkable. For comparison, most competitive forecasting models require dataset-specific training that takes hours to days. TimesFM generates predictions in seconds. 9,700+ GitHub stars as of early 2026, with 1,900 stars gained in a single week after Google announced integration into Connected Sheets. That integration is the business case: non-technical users can now run foundation model forecasts directly inside Google Sheets. The paper was published at ICML 2024. The codebase is Python, built on JAX and PAX. Checkpoints are available on the Hugging Face TimesFM collection in both JAX and PyTorch formats. If you work with time series data, this is the open-source model to know. For related AI tools, check data science tools on Skila. For more trending AI repositories, browse Skila Repos.