Google releases TimesFM forecasting model

Google Research has put a forecasting model family called TimesFM into the open, giving developers a public way to test foundation-model approaches on time-series data such as demand, traffic, sales and other sequential signals. Google’s latest research post, published on August 31, introduced TimesFM-3 as a “zero-shot foundation model for multivariate forecasting,” while the project’s public GitHub repository and Hugging Face pages show the broader model line has been available as open-source code and downloadable checkpoints. The release matters because forecasting tools are usually built for a single dataset, a single business problem or a narrow statistical setup. Google describes TimesFM instead as a pretrained model that can be applied to unseen datasets without task-specific retraining, a setup the company has repeatedly labeled “zero-shot” forecasting. ### So what exactly did Google release? Google Research said on August 31 that TimesFM-3 is a model for multivariate forecasting that can produce forecasts for multiple related time series in a single forward pass. (research.google) In the same post, Google said the model outperformed other forecasting models across major benchmarks. Google’s public repository describes TimesFM more broadly as a pretrained time-series foundation model for forecasting, and GitHub release pages show a current “TimesFM-3.0” release alongside earlier versions. (research.google) Hugging Face also lists public checkpoints across several generations, including 1.0, 2.0, 2.5 and 3.0 variants. ### What does “zero-shot” mean here? Google Research said its earlier TimesFM work was designed to forecast on unseen datasets without task-specific training. (research.google) In practical terms, that means a developer can feed a historical series into the model and ask for a forecast without first fitting a custom model for that exact dataset. Google’s later research on few-shot time-series learning framed that zero-shot capability as the baseline: the company said its “previous model, TimesFM, was a zero-shot, pre-trained foundation model” and then explored whether a few examples from related series could improve results further. (github.com) ### Where does the “100 billion data points” claim come from? Google Research said in its original TimesFM blog post that it trained the model on “100B real world time-points.” The company said most of that pretraining corpus came from Google Trends search-interest series and Wikipedia pageview data. (research.google) Google Cloud later described TimesFM in product posts as being pretrained on “over 400 billion real-world time-points,” which suggests Google expanded the training corpus in later versions. (research.google) That means the 100 billion figure is accurate for the original model family announcement, but not necessarily for the newest release. ### Can developers actually run it themselves? Google’s GitHub repository includes installation instructions, inference examples and release tags, and Hugging Face hosts downloadable checkpoints that can be run outside Google’s own products. (research.google) Those public artifacts support the claim that developers can use the model directly rather than only through a managed API. Google Cloud has also integrated TimesFM into BigQuery ML and AlloyDB workflows, according to product blog posts published in 2025. (cloud.google.com) Those posts describe AI.FORECAST and database integrations built around the same model family. ### What kinds of use cases does this cover? Google’s materials describe TimesFM as a general forecasting model rather than a tool for one vertical. The company’s examples and descriptions point to common time-series tasks such as predicting future values from historical sequences across business and operational settings. (github.com) The user claim that it can be used for sales, demand, prices, traffic or crypto volatility is directionally consistent with how foundation forecasting models are marketed, but Google’s primary materials frame the product more generally as a time-series forecaster rather than naming all of those categories in one official list. (cloud.google.com) ### Why would developers care beyond the model itself? GitHub and Hugging Face availability make TimesFM useful as a building block for forecasting pipelines, because developers can test inference, evaluation and deployment on public checkpoints instead of starting from scratch. (github.com) Google’s Cloud posts also show a path from research model to production-style tooling in SQL and database environments. As of September 8, 2026, the clearest next reference points are Google Research’s August 31 TimesFM-3 post, the `google-research/timesfm` repository releases, and the public Hugging Face model pages, where new checkpoints and usage notes are posted. (research.google) (github.com)