Google's TabFM: zero-shot tabular AI without training or tuning

Google Research has introduced TabFM (Tabular Foundation Model), a zero-shot model designed to handle classification and regression tasks on structured, tabular data with mixed numerical and categorical columns. Unlike traditional machine learning approaches that require extensive fine-tuning, hyperparameter optimization, and separate training pipelines, TabFM operates in a single forward pass. It takes training examples as part of its input context and outputs predictions directly — no separate training phase is needed. This makes it a true foundation model for tabular data, akin to how large language models perform zero-shot text tasks.

TabFM's architecture leverages transformer-based encodings adapted for mixed-type columns, allowing it to generalize across diverse tabular datasets without retraining. For professionals working with spreadsheets, databases, or CSV files, this means they can apply machine learning predictions instantly — simply feed in a few labeled examples and get answers. Potential use cases include quick baseline modeling, data exploration, and serving non-expert users who need ML on tabular data without coding complex pipelines. While still early-stage, TabFM signals a shift toward foundation models for structured data, similar to what we've seen in NLP and vision.