Google AI unveiled Tab FM, a zero-shot foundation model for tabular classification and regression, now embedded in BigQuery ML. The model treats a whole table as a prompt, using in-context learning to bypass training, hyperparameter tuning, and manual feature engineering. By eliminating these steps, data scientists can generate predictions for new datasets with a single forward pass.

Traditional pipelines rely on tree-based algorithms such as XGBoost or random forests, which demand extensive hyperparameter searches and domain-specific feature crafting. Tab FM's hybrid architecture combines alternating row‑column attention, row compression, and a transformer over compressed embeddings, enabling efficient processing of two-dimensional tables despite their orderless nature. The model was pre‑trained on hundreds of millions of synthetic tables generated via structural causal models.

Benchmarks on the Tab Arena suite show Tab FM outperforms tuned supervised baselines across 38 classification and 13 regression tasks, with an ensemble variant adding SVD features and Platt scaling for extra gain. The code and pretrained weights are released on Hugging Face and GitHub, and Google plans to expose the model through an AI.PREDICT SQL command in BigQuery, removing the need for ML expertise.