Google’s TimesFM v2.5 lands in BigQuery
What happened
Google Research released TimesFM v2.5, a 200‑million‑parameter foundation model for time‑series forecasting that supports 16k context windows and quantile predictions, and is now accessible via BigQuery for downstream use. The model’s quantile outputs and long context make it immediately useful for risk‑modeling and scenario forecasting without full retraining. (x.com)
Why it matters
Google Research has published TimesFM v2.5 and Google has hooked that release directly into BigQuery, so the pretrained forecasting model can be run inside BigQuery without pulling models into separate servers. (cloud.google.com) (docs.cloud.google.com) TimesFM v2.5 returns probabilistic forecasts as multiple percentiles (also called quantiles), so each forecast comes with a low, median, and high outcome instead of a single point estimate, and the model can attend to roughly sixteen thousand past observations when forming those forecasts — meaning much longer histories are available to the predictor than in prior versions. (huggingface.co) (github.com) Under the hood the checkpoint is compact: the main TimesFM 2.5 model uses about 200 million parameters, where each parameter is a number the model learned during pretraining to represent patterns in time series; the release also offers an optional quantile head of roughly 30 million parameters that produces those percentile outputs for forecasting horizons up to about 1,000 steps. (github.com) (letsdatascience.com) TimesFM is a decoder‑only transformer architecture that was pretrained on a very large corpus of time-stamped data (the original TimesFM training used billions of time points), which is why it can often produce usable forecasts for new series without dataset‑specific retraining — a behavior called “zero‑shot” forecasting. (research.google) (github.com) BigQuery exposes TimesFM as a native forecasting option inside BigQuery ML so inference runs on Google’s data platform (no external endpoints to host), the platform provides a SQL-accessible forecast function and recent Workspace updates added the ability to generate TimesFM forecasts from Google Sheets via Connected Sheets, and official checkpoints are also published on Hugging Face for local experimentation. (docs.cloud.google.com) (workspaceupdates.googleblog.com) (huggingface.co)
Key numbers
- Google Research released TimesFM v2.5, a 200‑million‑parameter foundation model for time‑series forecasting that supports 16k context windows and quantile predictions, and is now accessible via BigQuery for downstream use.
- (x.com) Google Research has published TimesFM v2.5 and Google has hooked that release directly into BigQuery, so the pretrained forecasting model can be run inside BigQuery without pulling models into separate servers.
Quick answers
What happened in Google’s TimesFM v2.5 lands in BigQuery?
Google Research released TimesFM v2.5, a 200‑million‑parameter foundation model for time‑series forecasting that supports 16k context windows and quantile predictions, and is now accessible via BigQuery for downstream use. The model’s quantile outputs and long context make it immediately useful for risk‑modeling and scenario forecasting without full retraining. (x.com)
Why does Google’s TimesFM v2.5 lands in BigQuery matter?
Google Research has published TimesFM v2.5 and Google has hooked that release directly into BigQuery, so the pretrained forecasting model can be run inside BigQuery without pulling models into separate servers. (cloud.google.com) (docs.cloud.google.com) TimesFM v2.5 returns probabilistic forecasts as multiple percentiles (also called quantiles), so each forecast comes with a low, median, and high outcome instead of a single point estimate, and the model can attend to roughly sixteen thousand past observations when forming those forecasts — meaning much longer histories are available to the predictor than in prior versions. (huggingface.co) (github.com) Under the hood the checkpoint is compact: the main TimesFM 2.5 model uses about 200 million parameters, where each parameter is a number the model learned during pretraining to represent patterns in time series; the release also offers an optional quantile head of roughly 30 million parameters that produces those percentile outputs for forecasting horizons up to about 1,000 steps. (github.com) (letsdatascience.com) TimesFM is a decoder‑only transformer architecture that was pretrained on a very large corpus of time-stamped data (the original TimesFM training used billions of time points), which is why it can often produce usable forecasts for new series without dataset‑specific retraining — a behavior called “zero‑shot” forecasting. (research.google) (github.com) BigQuery exposes TimesFM as a native forecasting option inside BigQuery ML so inference runs on Google’s data platform (no external endpoints to host), the platform provides a SQL-accessible forecast function and recent Workspace updates added the ability to generate TimesFM forecasts from Google Sheets via Connected Sheets, and official checkpoints are also published on Hugging Face for local experimentation. (docs.cloud.google.com) (workspaceupdates.googleblog.com) (huggingface.co)
