Introducing TabFM: A zero-shot foundation model for tabular data

TabFM is a new foundation model designed to handle classification and regression tasks on tabular data, the structured data format that powers most enterprise predictive applications. Rather than requiring the traditional process of manual hyperparameter tuning and feature engineering that tree-based algorithms like XGBoost demand, TabFM applies a "zero-shot" approach similar to how large language models learn new tasks from examples provided in context without updating underlying model weights. The model frames tabular prediction as an in-context learning problem, processing entire datasets as unified inputs to generate predictions on new data in a single forward pass. This approach addresses a longstanding bottleneck in machine learning workflows where data scientists must invest significant effort preparing and optimizing models for each new dataset.