Introducing TabFM — a zero-shot foundation model for tabular data

Abstraction: Google foundation model does zero-shot tabular prediction

Key points:

• Google Research (Kong & Das, June 2026) introduces TabFM: a foundation model for tabular classification/regression that frames prediction as in-context learning — the whole table (train rows + test rows) goes in as one prompt, predictions come out in a single forward pass. No .fit(), no hyperparameter tuning, no feature engineering.
• Aims to displace the XGBoost/random-forest workflow that has dominated tabular ML; follows the same "zero-shot" playbook as Google's TimesFM did for time-series.
• Architecture is a hybrid synthesizing TabPFN and TabICL ideas, handling tables' 2-D, order-invariant structure (row/column swaps don't change meaning) that vanilla LLM tokenization can't.
• Trained entirely on hundreds of millions of synthetic datasets generated via structural causal models (SCMs) — because large, diverse real industrial tables are proprietary/scarce.
• Evaluated on TabArena (Elo, head-to-head): 38 classification + 13 regression datasets, 700–150k samples. Weights open on Hugging Face + GitHub.

Connections: Google · Tabfm · Hugging Face · Xgboost · Foundation Models · In Context Learning · Tabular Machine Learning · Synthetic Data

Source: https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/