Google Adds TabFM to BigQuery, Enabling 20-Feature Predictions Without Model Training
Updated
Updated · InfoWorld · Sep 2
Google Adds TabFM to BigQuery, Enabling 20-Feature Predictions Without Model Training
3 articles · Updated · InfoWorld · Sep 2
Summary
Google has put its TabFM tabular foundation model into BigQuery preview, letting users run classification and regression predictions from structured data through AI.PREDICT instead of training and deploying separate ML models.
Using labeled historical rows as in-context examples, TabFM can handle categorical fields and missing values; Google also added AI.EVALUATE so teams can measure prediction accuracy against known outcomes.
For enterprises, that could let SQL users generate forecasts inside BigQuery without moving data to another ML platform, cutting engineering work, infrastructure overhead and some governance complexity.
Google still steers customers to XGBoost or custom models for very large datasets, explainability needs or workloads exceeding TabFM's current 20-feature limit, making it more a targeted tool than a full replacement.
Starting Oct. 30, 2026, TabFM will shift to token-based pricing on top of normal BigQuery charges, which analysts said could erode cost advantages for high-volume prediction workloads.