Google Research releases TabFM, a zero-shot foundation model for tabular data

Google Research has unveiled TabFM, a foundation model for tabular data that can perform zero-shot classification and regression without training on a specific dataset, MarkTechPost reported on July 1, 2026.

How TabFM processes tables

TabFM produces predictions in a single forward pass using in-context learning: the model draws on examples inside the prompt itself on the fly, with no separate training loop for each dataset.

• Format: foundation model for tabular data
• Capabilities: zero-shot classification and regression
• Method: in-context learning, prediction in a single forward pass
• Does not require: dataset-specific training, hyperparameter tuning, manual feature engineering
• Architecture: hybrid attention

The model's name points to a hybrid attention architecture. It lets the model account for relationships between rows (objects) and between columns (features) at the same time, something classic text-oriented transformers don't do by default.

Why tabular data is a distinct problem for AI

Most enterprise data isn't text or images but tables: rows of customers, transactions, and metrics. For a long time, the best tools for this work have been gradient boosting and classic ML models such as XGBoost and LightGBM. They're accurate, but each one has to be retrained and tuned for a specific dataset and a specific task, and any change to the feature set means another round of data preparation.

The idea behind a tabular foundation model is that one model can serve many tasks: pretrain a single large model on a wide range of heterogeneous tables, then apply it to new data without retraining. It mirrors how language models handle new text tasks without fine-tuning when prompted correctly. For tabular data, this approach was long considered hard to achieve precisely because tables are so heterogeneous: each has its own set of columns, data types, and semantic relationships between them. Text is different, since all documents share the same grammar and vocabulary.

What this means for data scientists

If TabFM's zero-shot approach proves accurate enough on real enterprise data, it changes the economics of routine ML tasks on tables. The "collect a dataset → train a model → tune hyperparameters → deploy" cycle would shrink to a single step: feed the data into the model and get a prediction. That's especially valuable for tasks where the dataset is small or changes quickly, making it uneconomical to train a dedicated model.

What it means

Google Research is once again carrying the foundation-model logic of "train once, apply everywhere" from text and images into tabular data, a domain that has so far remained the preserve of classic machine learning. If the approach holds up in practice across a wide range of tasks, competition will play out not only among different companies' tabular foundation models but also between this new class of models and familiar gradient boosting, over which becomes the default tool for routine ML tasks on tabular data.