Tabular Foundation Overview
Models, labs, and the evolving landscape
Tabular Foundation Models
The pre-trained models, the labs, the research.
At a Glance
The Space
Tabular foundation models make predictions on new datasets without a classic training step: no hyperparameter tuning, no gradient descent at inference time. They are pre-trained on large collections of tabular data (often synthetic) and generalize to new tables in a single forward pass.
The field is young but moving fast. Open-source models like TabPFN and TabICL match tuned gradient-boosted trees on standard benchmarks. Startups like Prior Labs and Fundamental have raised significant rounds to commercialize these approaches, while Amazon, Microsoft, and SAP are shipping tabular FMs inside existing platforms.
About Me
I’m Christoph Molnar, a machine learning author, educator, and consultant. I’ve written several books on topics like interpretable machine learning and conformal prediction. I’m currently exploring the tabular foundation model space and am excited about how fast things are moving. This site is my way of tracking the models, the labs, and the landscape.