TabFM: Google's foundation model for tabular data

By TheSequence.

The email opens with a common joke in the machine learning world: that the most valuable model in the field isn't a transformer, but a set of gradient boosting trees (XGBoost-style) trained on a CSV. According to the author, despite all the media noise around frontier models, the real workhorses of enterprise ML are still XGBoost pipelines predicting churn, fraud and credit risk over rows and columns. The workflow around these models has barely changed in a decade: load the table, engineer features (feature engineering), cross-validate, tune hyperparameters, and repeat the ritual with every new dataset.

The article notes that Google Research has just taken direct aim at that ritual with TabFM, a model released a few days ago. It is described as a foundation model for tabular classification and regression capable of generating predictions on tables it has never seen before, in a single forward pass, with no need for training, hyperparameter tuning or feature engineering. The way it works is to hand it the entire problem at once —training rows, test rows, everything— as a single giant prompt, and the model responds. This is, according to the text, in-context learning applied to spreadsheets.

The email points out that this approach is not an isolated novelty from the team: TabFM would follow the same 'playbook' as TimesFM, the time-series foundation model from the same Google team, which would have quietly become one of the most deployed research artifacts the company has released. The author suggests that to properly understand TabFM you first need to understand that TimesFM heritage, and the email cuts off just as it begins to develop that background ('The TimesFM prelude...'), so no further technical detail about the architecture, benchmarks or concrete TabFM results is available in the received body.

Before the technical content, the newsletter itself includes a note from the author explaining that TheSequence has been running for more than two years without sponsors, despite receiving constant sponsorship offers, as a way to maintain its technical independence and objectivity, and invites readers to subscribe (free or paid) as a way to support it.