TabFM zero-shot foundation model for tabular data

Category: ML research
Status: developing
Sources: Google Research
Summary: Google published TabFM on 2026-06-30, a foundation model for tabular classification and regression that predicts on unseen tables in a single forward pass without task-specific training, tuning, or feature engineering. The architecture alternates row and column attention, compresses rows into dense vectors, and applies a transformer in-context-learning layer; it was trained on hundreds of millions of synthetic datasets from structural causal models. Google reports evaluation on TabArena (38 classification and 13 regression datasets) with base and ensemble configurations, and says weights and code are on Hugging Face and GitHub with BigQuery integration planned.
Why it matters: Zero-shot tabular prediction targets the most common enterprise data shape, where gradient-boosted trees and manual feature work still dominate.
Follow-up: Watch for independent TabArena reproduction and comparisons against tuned gradient-boosting baselines.