This page may contain stale information. Last updated: 2026-07-01
Overview
TabFM (Tabular Foundation Model) is google Research’s zero-shot tabular ML model released June 30, 2026. Performs classification and regression via in-context learning in a single forward pass — no per-dataset training or hyperparameter tuning.
Architecture
- Hybrid attention: TabPFN-style row/column attention + TabICL compression
- Pre-trained on hundreds of millions of synthetic structural causal model (SCM) datasets
- TabFM-Ensemble: cross/SVD features, 32-way ensemble, Platt scaling for classification
Distribution
- pip:
tabfmv1.0.0 (PyPI, June 30, 2026) - GitHub: google-research/tabfm (Apache 2.0)
- Hugging Face: google/tabfm-1.0.0-pytorch (non-commercial model license)
- BigQuery:
AI.PREDICTSQL integration planned (google-cloud)
GitHub disclaimer: not an officially supported Google product.
Use Cases
Enterprise tabular ML: fraud detection, churn prediction, risk scoring — domains where XGBoost/lightGBM traditionally dominate.