Google launches TabFM to bring zero-shot foundation models to tabular data

TabFM is a zero-shot foundation model from Google Research built specifically for tabular regression and classification.

Traditional tabular machine learning is notorious for its manual overhead. Getting reliable results out of tree models like XGBoost usually means hours spent on feature engineering and hyperparameter tuning. TabFM bypasses that entire loop by applying in-context learning to tabular data, generating predictions on unseen tables in a single forward pass without updating model weights.

How it works under the hood

Language models process one-dimensional ordered sequences, but tabular data is inherently two-dimensional and orderless: reordering rows or columns does not alter the underlying data. To solve this, TabFM combines design elements from TabPFN and TabICL into a hybrid attention architecture.

During inference, the model treats the whole dataset—both historical training examples and target test rows—as a single unified prompt. It maps cross-row and cross-column interactions on the fly without any dataset-specific gradient updates.

Because real-world enterprise tables are proprietary and scarce in public domains, TabFM was pre-trained entirely on hundreds of millions of synthetic datasets. Google used structural causal models to generate vast synthetic distributions, allowing the network to capture complex feature relationships that generalize well to real-world datasets.

Performance and where to try it

Google evaluated TabFM on the TabArena benchmark across 38 classification and 13 regression datasets, with table sizes ranging from 700 to 150,000 samples. TabFM consistently outperformed heavily tuned supervised baselines like XGBoost and random forests out of the box.

TabFM is available now on Hugging Face and GitHub. For enterprise workflows, Google has natively integrated TabFM into Google Cloud BigQuery, allowing users to run zero-shot predictions directly via standard SQL queries using the AI.PREDICT function.

Why it matters

For data teams, TabFM turns what used to be a multi-day modeling sprint into a single SQL query. It is ideal for baseline evaluations, quick prototyping, and fast classification tasks where manual feature engineering simply takes too long.

Tree-based algorithms have held tabular data in a chokehold for over a decade because deep learning rarely justified the extra setup. TabFM shows that in-context learning on massive synthetic data can finally make tabular foundation models both genuinely competitive and radically easier to deploy.