Enterprise predictive analytics typically requires building custom models with libraries like XGBoost, involving complex training, tuning, and deployment cycles. TabFM, developed by Google Research, is a pre-trained foundation model for tabular regression and classification that uses in-context learning. It is integrated natively into BigQuery and delivers zero-shot predictions via a single SQL query using the AI.PREDICT and AI.EVALUATE functions, eliminating separate training and deployment steps. The model automatically handles featurization, missing values, and categorical encoding.

On the TabArena benchmark, TabFM consistently outperforms both classic ML models and other tabular foundation models in classification and regression. Under the hood, BigQuery performs distributed parallel inference with intelligent training-data sampling to efficiently process large datasets. TabFM is best suited for small-to-medium historical datasets, frequently changing data, and use cases needing rapid predictive insights without ML expertise, including conversational or agentic workflows. Traditional models like XGBoost remain preferable for very large datasets, when full control over hyperparameter tuning is needed, for high feature counts, or when feature importance explainability is required.