Google’s TabFM entered BigQuery preview, providing zero-shot tabular regression and classification through SQL across millions of inference rows.

Google introduced TabFM in BigQuery as a preview service. Developed by Google Research, TabFM is a pretrained foundation model for regression and classification using tabular data.

TabFM uses in-context learning and can perform zero-shot predictions without a separate model-training, tuning or artifact-deployment process. Users provide labeled historical data and new prediction tables through a Structured Query Language statement.

The model is available through the built-in `AI.PREDICT` and `AI.EVALUATE` SQL functions. It automatically determines whether a task is classification or regression from the target label’s data type and handles preprocessing tasks including missing values and categorical encoding.

Google says BigQuery’s distributed inference architecture can process inference tables containing up to millions of rows in minutes. Google also reported that TabFM outperformed classic machine-learning systems and other tabular foundation models in its TabArena classification and regression evaluations.