Google Research has introduced TabFM, a hybrid-attention foundation model designed for tabular classification and regression. The model enables zero-shot predictions by processing entire datasets as a single prompt in one forward pass. This approach eliminates the need for per-dataset training, hyperparameter tuning, and manual feature engineering.