Google Research Releases TabFM — Foundation Model for Tabular Data with Zero-Shot

Google Research released TabFM — foundation model for tabular data with zero-shot

Google Research presented TabFM — a foundation model for tabular data capable of zero-shot classification and regression without training on specific datasets; this was reported by MarkTechPost portal on July 1, 2026.

How TabFM processes tables

TabFM gets predictions in a single forward pass, using in-context learning — the model "on the fly" relies on examples within the request itself, without a separate training cycle for each dataset.

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

The model's name declares a hybrid attention architecture — it allows simultaneously accounting for relationships between rows (objects) and between columns (features) of a table, which in classical text-focused transformers is not built in by default.

Why tabular data is a separate problem for AI

Most corporate data is not text or images, but tables: customer rows, transactions, metrics. For a long time the best tools for such tasks remained gradient boosting and classical ML models like XGBoost and LightGBM — they are accurate, but each must be retrained and tuned for a specific dataset and specific task, and any change in feature composition means a new data preparation cycle.

The idea of a foundation model for tables is for one model to work for many tasks: train one large model on many diverse tables in advance, then apply it to new data without retraining — by analogy with how language models solve new text tasks without fine-tuning if asked correctly. For tabular data this path was long considered difficult to achieve precisely because of the heterogeneity of tables: each has its own set of columns, data types and semantic relationships between them, unlike text where grammar and vocabulary are common to all documents.

What this means for data scientists

If TabFM's zero-shot approach works well enough on real corporate data, it changes the economics of typical ML tasks with tables: instead of the cycle "collect dataset → train model → tune hyperparameters → deploy" there will be just one step — feed data to the model and get a prediction. This is especially valuable for tasks where the dataset is small or changes quickly, and it's not cost-effective to train a separate model for it.

What this means

Google Research once again transfers the logic of foundation models — "train once, apply everywhere" — from text and images to the field of tabular data, which until now has remained the domain of classical machine learning. If the approach is confirmed in practice across a wide range of tasks, competition will unfold not only between foundation models for tables from different companies, but also between this new class of models and familiar gradient boosting — over who will become the default tool for typical ML tasks with tabular data.