Zero‑Shot Tabular Predictions with TabFM in BigQuery: What Engineers Need to Know

Google has introduced TabFM, a pre‑trained foundation model for tabular regression and classification that lives inside BigQuery. Engineers can now request predictions with a single AI.PREDICT SQL call, eliminating the separate training, tuning, and deployment steps that traditionally dominate predictive pipelines.

How TabFM Changes the Predictive Workflow

TabFM operates in a zero‑shot mode: you supply a historical table containing the target column and a separate table with new rows, and the model infers the task type automatically. The built‑in AI.PREDICT function handles missing values, categorical encoding, and other featurization steps, so there is no need to construct explicit feature pipelines.

SELECT *
FROM AI.PREDICT(
  TABLE `my_project.my_dataset.historical_transactions`,
  TABLE `my_project.my_dataset.new_transactions`,
  label_col => 'is_fraud'
);

For model assessment, AI.EVALUATE runs the same in‑context evaluation against a held‑out set, providing standard metrics without exporting data.

Operational Considerations

The model runs natively on BigQuery’s distributed inference engine, allowing inference over millions of rows in minutes. Because the service is in preview, availability is limited to the preview environment, and engineers should monitor usage quotas and latency as part of their deployment checklist. The SQL‑centric interface means existing CI/CD pipelines that manage BigQuery queries can incorporate predictive steps without adding separate ML infrastructure.

Security and Access Implications

All data used for training and inference remains inside BigQuery, so the platform’s existing IAM policies continue to govern who can read the source tables and invoke AI.PREDICT. Practitioners should verify that only authorized principals have SELECT access to the training and prediction tables, and consider using dataset‑level permissions to enforce least‑privilege access. Because the model does not export data outside the warehouse, the attack surface is limited to the usual BigQuery access controls.

Evaluating Model Fit

TabFM’s claim of “state‑of‑the‑art accuracy” is backed by its performance on the TabArena benchmark, where it reportedly outperforms classic ML models and other tabular foundation models. Teams should compare TabFM’s evaluation scores against any existing custom models on their own validation sets to confirm the expected gain before fully replacing legacy pipelines.

Related CloudNinjas coverage: Google Cloud.

What This Means For Practitioners

• Zero‑shot SQL predictions reduce the time and expertise needed to operationalize new models.
• Existing BigQuery CI/CD and monitoring frameworks can be extended to cover predictive workloads.
• Maintain strict IAM controls on the tables used for in‑context examples to prevent unintended data exposure.
• Use AI.EVALUATE to benchmark TabFM against current models before decommissioning legacy training pipelines.
• Plan for the preview status by testing in a non‑production project and monitoring quota usage.