Google AI’s TabFM Redefines Zero-Shot Tabular Predictions

Google AI rolled out TabFM on July 1, 2026, a foundation model that predicts tabular data without training. It handles classification and regression on unseen datasets with zero tuning or feature engineering.

TabFM blends row and column attention inspired by TabPFN with in-context learning from TabICL. This hybrid approach lets it learn relationships within tables dynamically, rather than relying on dataset-specific adjustments.

Training involved hundreds of millions of synthetic datasets created on the fly using structural causal models. This extensive, synthetic data exposure equipped TabFM to generalize across diverse real-world tables.

TabFM was evaluated on TabArena, a benchmark measuring Elo scores from head-to-head matches across 38 classification and 13 regression datasets. It outperformed heavily tuned industry-standard supervised models consistently.

There are two versions: a base model running a single forward pass without tuning and TabFM-Ensemble. The ensemble adds cross features, SVD components, a 32-way ensemble, non-negative least squares weighting, and Platt scaling.

Google plans to integrate TabFM into BigQuery through an `AI.PREDICT` SQL command. This will let analysts apply zero-shot predictions directly within a familiar data warehouse environment.

TabFM is open source and available on Hugging Face and GitHub. Google Research did not name any outside investors or funding sources for the project.

The model’s spatial reasoning draws on geography principles. “The first law of geography is that ‘everything is related to everything else, but near things are more related than distant things’,” said Dr. Mingshu Wang.

Rui Deng explained, “We divide the region into a grid, understanding relative distances between data points. Then, we guide the model to attend more to nearer points, focusing on the local context.”

Dr. Ziqi Li noted, “This study demonstrates how established geographical principles can be seamlessly integrated into a pre-trained foundation model, improving its spatial understanding and handling larger datasets.”

The arrival of TabFM coincided with several Google Cloud announcements. BigQuery Conversational Analytics launched, allowing natural language queries within BigQuery.

Google Cloud also introduced the Open Knowledge Format, a vendor-neutral standard for AI metadata and curated knowledge. This aims to improve AI interoperability and context sharing.

Collaborations with Apple produced Private Cloud Compute systems using Confidential Computing and Titanium security architecture. Anthropic’s Claude Fable 5 model became generally available on Google Cloud.

Cloud Atelier, a virtual shopping experience, debuted alongside AI security highlights. These included embedding AI agents into software development and guidance from Mandiant on AI threat defenses.

Google Cloud finalized its acquisition of Wiz, a cloud and AI security platform. Its AI product line now centers on the Gemini Enterprise Agent Platform, with tools like Agent Designer, long-running agents, and project-based memory controls.

Infrastructure and model updates rolled out as well, including Gemini 3.5 Flash, Gemini Omni, Nano Banana 2 and Pro, Gemini Embedding 2, and Veo 3.1 Lite. The focus is clear: standards, managed agents, security controls, and enterprise workflow integration.

Suzie Millar, commenting on BigQuery Conversational Analytics, said, “Analysis that used to take weeks can now be done in minutes, saving our financial analysts around half a day each week.”

She added, “By making analysis more self-serve, we’re helping teams create faster insight to support better product and commercial decision-making.”