Google open‑sources TimesFM
Google open‑sourced TimesFM, a pre‑trained time‑series forecaster that the announcement claims outperforms supervised baselines in zero‑shot settings and can be used for predictions without fine‑tuning. The release positions TimesFM as a plug‑in forecasting model for enterprise prediction tasks. (x.com)
Time-series forecasting is the business of predicting the next numbers in a sequence, from power demand to retail sales. Google has released TimesFM as open-source code and model checkpoints through its Google Research repository and Hugging Face collection. (github.com) Google Research describes TimesFM as a pre-trained forecasting model, meaning developers can feed it past values and generate forecasts without training a custom model first. The project page lists TimesFM 2.5 as the latest open version and says archived 1.0 and 2.0 releases remain available. (github.com) The underlying paper, “A decoder-only foundation model for time-series forecasting,” was presented at the International Conference on Machine Learning in 2024. Google’s research blog said the model was pre-trained on a corpus of 100 billion real-world time points and was built to work across datasets with different domains and time granularities. (arxiv.org, research.google) Google’s claim is that this pre-training lets TimesFM make “zero-shot” forecasts, which means using a new dataset without additional task-specific tuning. In the paper abstract, the authors said its out-of-the-box accuracy on public datasets came close to state-of-the-art supervised models trained separately for each dataset. (arxiv.org) The model borrows an idea from language models: break a long sequence into chunks, then predict the next chunk from the earlier ones. Hugging Face’s documentation says TimesFM is a decoder-only transformer that uses non-overlapping patches of time-series data and generates forecasts autoregressively. (huggingface.co) Google is also pushing TimesFM beyond research code and into products. BigQuery documentation published in April 2026 says its built-in TimesFM univariate model is an implementation of the open-source Google Research model and is available in all BigQuery-supported regions. (cloud.google.com) That product tie-in puts the open release in the middle of a broader race to turn “foundation models” into standard tools for forecasting, not just text generation. Google’s repository describes TimesFM as a plug-in model for time-series forecasting, and its BigQuery documentation frames it as a model that can be applied to new forecasting datasets across many domains. (github.com, cloud.google.com) The open repository also shows how Google has been iterating on the model after the first public release. The current README says TimesFM 2.5 is a 200 million-parameter release, while the Hugging Face collection lists earlier 1.0 and 2.0 checkpoints, including 500 million-parameter versions updated in 2025. (github.com, huggingface.co) Google’s own materials include caveats alongside the release. The GitHub repository and model cards say the open version is “not an officially supported Google product,” even as Google Cloud offers a supported implementation inside BigQuery. (github.com, huggingface.co) For companies that forecast inventory, traffic, prices, or demand, the pitch is straightforward: use one pre-trained model instead of building a separate forecasting system for every series. Google’s release turns that pitch into downloadable code, published weights, and a cloud product that now points to the same model family. (github.com, cloud.google.com)
