Google Research introduced TabFM June 30, 2026 — foundation model for tabular classification and regression via in-context learning. Single forward pass; no per-dataset training, hyperparameter tuning, or manual feature engineering.
Hybrid attention architecture; trained on hundreds of millions of synthetic structural causal model datasets. Available Hugging Face and GitHub. BigQuery AI.PREDICT SQL integration coming in coming weeks.
TabFM reframes tabular prediction as ICL problem; outperforms heavily tuned supervised algorithms in research benchmarks. TabFM-Ensemble variant adds cross/SVD features and 32-way ensemble weighting.
Disclaimer on GitHub: not officially supported Google product.