Google targets tabular AI with new foundation model TabFM

Google Research just dropped a model that could make a significant chunk of data science busywork obsolete. TabFM, announced on June 30 by researchers Weihao Kong and Abhimanyu Das, is a foundation model built specifically for tabular data, the kind of structured rows-and-columns information that powers everything from CRM systems to crypto exchange ledgers.

Here’s the thing: the vast majority of business data lives in tables, yet every new prediction task has traditionally required training a fresh model from scratch. TabFM skips that entirely. It treats tabular prediction as an in-context learning problem, generating classifications and regressions on datasets it has never seen before in a single forward pass. No training. No hyperparameter tuning. No feature engineering.

How TabFM actually works

In technical terms, the model uses training rows from a dataset as context within a single forward pass to make predictions on new rows. This is the same in-context learning paradigm that Google originally introduced with TimesFM for time-series forecasting, now adapted for the broader universe of structured data.

The model draws on lessons from prior work like TabPFN, but extends the approach into a fully open-source foundation model. It’s available right now on Hugging Face under google/tabfm-1.0.0-pytorch and on GitHub at google-research/tabfm, distributed under the Apache-2.0 license.

Why crypto should pay attention

Tabular data is the backbone of the crypto industry, even if nobody talks about it that way. Trading records, on-chain transaction logs, DeFi protocol metrics, compliance datasets, risk models. All tables. All currently requiring bespoke models whenever a fund or exchange wants to extract predictions from them.

The caveats that actually matter

Explainability is the first concern. In high-stakes financial applications, regulators and risk managers want to understand why a model made a specific prediction. Foundation models, by their nature, are less transparent than purpose-built models with carefully engineered features.

Then there’s the question of performance under extreme distribution scenarios. Financial data, and crypto data in particular, is notorious for heavy-tailed distributions. Whether TabFM handles these edge cases gracefully is an open question that will require rigorous validation before anyone stakes real capital on its outputs.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.