Google AI Introduces TabFM
July 1, 2026
Google Research introduced TabFM, a foundation model built for tabular data. TabFM performs classification and regression without dataset-specific training. Every prediction comes from a single forward pass. The model reframes tabular prediction as an in-context learning problem. It is available now on Hugging Face and GitHub.
Key capabilities
- TabFM predicts on unseen tables with no training, tuning, or feature engineering.
- It reads the full dataset as one prompt, then predicts via in-context learning.
- The architecture combines TabPFN-style row/column attention with TabICL-style in-context learning.
- Training used hundreds of millions of synthetic datasets from structural causal models.
- Google BigQuery will expose TabFM through an AI.PREDICT SQL command soon.
Tabular data forms the backbone of enterprise data infrastructure — customer churn, fraud detection, and similar tasks live in tables. Tree-based methods (XGBoost, random forests) dominated but required extensive hyperparameter tuning and feature engineering per dataset. TabFM applies zero-shot logic from LLMs to tables.
Architecture
Traditional models update parameters per dataset. TabFM takes the whole dataset as a unified prompt holding training examples and target testing rows.
Three mechanisms:
- Alternating row and column attention: Following TabPFN, attention alternates across columns (features) and rows (examples).
- Row compression: Each row’s cross-attended information compresses into a single dense vector.
- In-context learning: A Transformer runs over compressed embeddings following TabICL.
Training
TabFM trains entirely on hundreds of millions of synthetic datasets generated dynamically using structural causal models (SCMs). Google reports strong generalization to unseen real-world data.
Benchmarking
Evaluated on TabArena: 38 classification and 13 regression datasets, sample sizes 700–150,000.
Two configurations: plain TabFM (single forward pass, no tuning) and TabFM-Ensemble (cross features, SVD features, 32-way ensemble with NNLS weights, Platt scaling for classification).
TabFM consistently outperforms heavily tuned industry-standard supervised algorithms per Google.
Getting started
Requires Python 3.11+, jax0.10.1, flax0.12.7. Pre-trained weights download from Hugging Face. scikit-learn compatible via TabFMClassifier and TabFMRegressor. fit() prepares encoders/scalers but does not train model weights.
Use cases: customer churn scoring, credit risk classification, house price regression — all without per-dataset training cycles.