Tabular Foundation Model

A great deal of data at companies and government agencies resides in tables: rows for customers, columns for age, revenue, or place of residence. Until now, a separate computational procedure was built for each such table, learning only from exactly that data. A Tabular Foundation Model takes a different approach. It was pretrained beforehand on millions of artificial or collected tables, learning general patterns of how columns relate to one another. Afterward, it is presented with an entirely new table and fills in missing values or predicts future values. The term “Foundation Model” means: a large base model that is reused for many different tasks instead of being built from scratch each time.

Why tables are the economy’s most important data

The well-known AI successes initially concerned text and images. But the everyday work of banks, insurers, and hospitals consists of tables. Whether a loan will be repaid, whether a machine is about to fail, which patient is at high risk: all of this is contained in rows and columns. Whoever predicts better here saves money directly.

Until now, this was effort for experts. A data analyst had to prepare columns, select a method, and painstakingly optimize its settings. This process took days to weeks per project. A pretrained model can greatly shorten this step because it delivers usable predictions immediately.

This is especially valuable with small amounts of data. A hospital may have only 400 patient records for a rare disease. Classical methods learn little that is reliable from this. A model that has already seen many similar data structures copes significantly better with such small tables.

Pretraining on invented tables

The central trick lies in pretraining. Developers generate huge quantities of artificial tables by computer, with known relationships. In one table, the target column depends roughly linearly on two others; in the next, it jumps abruptly; in the third, hardly at all. The model practices on millions of such cases, guessing the relationship from just a few rows. So it doesn’t learn facts about customers, but rather the craft of predicting.

When applied, the model is given the existing rows along with their known answers, plus the new row whose value is missing. The model compares, recognizes the pattern, and outputs a prediction. Additional training is often unnecessary altogether. Experts call this in-context learning: learning solely from the examples contained in the input window.

A comparison helps. A classical procedure is like a student who has to relearn the entire material from scratch for every exam. The Foundation Model is like an experienced teacher: he has seen thousands of task types and immediately recognizes the principle upon looking at new numbers. The limit lies in size. Many of these models work reliably only up to a few thousand rows and columns, because everything has to fit into the input at once.

TabPFN, cloud services, and footnotes in quarterly reports

The best-known example is called TabPFN and comes from research at the University of Freiburg. A version of it appeared in the journal Nature in 2025, which drew attention across the industry. Google, Amazon, and several start-ups are working on comparable models. In news reports, they often appear under buzzwords such as “Foundation Model for enterprise data.”

In everyday life, one usually notices nothing of this directly. One encounters the results: a credit decision, a bank’s fraud warning, a maintenance recommendation for a machine. Analytics software at companies is also increasingly building in such models. The user uploads a table and receives predictions without configuring anything themselves.

A common misconception: a Tabular Foundation Model is not a language model like ChatGPT. It doesn’t write text and doesn’t understand questions in everyday language. It processes numbers and categories within a fixed grid. That is why it often clearly outperforms language models at tabular tasks, despite being much smaller.