Google’s TabFM brings zero-shot prediction to business tables

Google Research has introduced TabFM, a foundation model designed to predict on unseen tabular datasets without per-dataset training, feature engineering, hyperparameter tuning or retraining pipelines. Instead of updating weights, users pass historical rows with labels and target rows together as context, allowing the model to infer relationships at runtime and return predictions in a forward pass.

TabFM avoids the weaknesses of general LLMs on tables by preserving the grid structure rather than serializing data into text. Its architecture combines alternating row and column attention, row compression and in-context learning, drawing on ideas from TabPFN and TabICL. Google trained the model entirely on hundreds of millions of synthetic datasets generated with structural causal models, aiming to learn tabular priors without using confidential real-world CSV files.

Benchmarks on TabArena covered 51 diverse tabular datasets across 38 classification and 13 regression tasks, where TabFM matched or beat heavily tuned supervised baselines in zero-shot mode. A TabFM-Ensemble setup runs 32 variations and blends the outputs for stronger results. Current limits include 10 output classes, optimization for up to 500 features, heavier inference, and non-commercial pretrained weights. Google is also integrating TabFM into BigQuery through an “AI.PREDICT” command, with the model best suited to rapid prototyping, data drift and small to medium-sized datasets under 100,000 rows rather than ultra-low-latency APIs or massive tables exceeding one million rows.