Foundation models have transformed text, images and time-series forecasting, but tabular data has remained dominated by traditional machine learning algorithms that must be trained separately for every new dataset.

Google is seeking to change that with TabFM, a foundation model that brings zero-shot, in-context learning to structured enterprise data, allowing users to generate predictions without manual model development.

The company said the model is intended for enterprise use cases ranging from fraud detection and customer analytics to financial forecasting and other structured data applications that have historically relied on algorithms such as XGBoost and random forests.

Instead of fitting a model to each new dataset, TabFM treats an entire table as context for in-context learning. Historical examples and unlabeled target rows are processed together, enabling the pretrained model to learn task-specific relationships during inference rather than through additional training. This allows predictions to be generated immediately in a single forward pass on previously unseen datasets.

The model uses a hybrid architecture that combines alternating attention across rows and columns with compressed row embeddings before applying transformer-based inference.

Google said this design captures complex feature interactions while reducing computational costs compared with applying attention directly across the entire table.

To overcome the shortage of large public tabular datasets, TabFM was pretrained exclusively on hundreds of millions of synthetic datasets generated using structural causal models.

Google said benchmark results on the TabArena evaluation suite showed TabFM consistently outperforming leading supervised machine learning models, while an ensemble version incorporating additional engineered features achieved even stronger performance.

The company plans to integrate TabFM directly into BigQuery, enabling users to run regression and classification models using the AI.PREDICT SQL function without specialized machine learning knowledge.