Tabular Foundation Model

A great deal of data in companies exists in the form of tables: rows for individual cases, columns for individual features. A bank has one customer per row and columns for age, income, or loan amount. Anyone wanting to derive predictions from this previously had to train their own computational procedure from scratch for each table. A tabular foundation model is an already fully prepared AI model that you simply feed with a new table. It has previously practiced on millions of artificially generated tables to learn what relationships between columns typically look like. This means it can immediately predict a column it hasn’t been shown, for example: Will this customer repay the loan?

The last bastion of classical methods

With text and images, large pre-trained models have long since displaced the older technology. With tables, this was different for a long time. There, so-called decision tree methods prevailed, which break down a table through many yes-no questions. These methods are fast, robust, and require little data. Precisely for this reason, tables were considered the area where modern AI models offered no advantage.

Since around 2023, this has turned around. Models such as TabPFN show that a single pre-trained model performs just as well or better on small tables. The practical benefit is above all time. Instead of optimizing settings for days, you get a usable result within seconds.

Economically, this is relevant because tabular data shapes everyday business. Accounting, inventory levels, patient records, sensor readings from machines: all of it is tables. If predictions on this suddenly become possible without a team of specialists, the entry barrier for smaller companies drops significantly.

Trained on invented tables

The decisive trick lies in the pre-training. You cannot feed a model half the internet as you would with text, because tables are usually confidential and completely different from one another. Instead, millions of artificial tables are generated by computer. In each case, a rule is randomly determined according to which the columns are related, and rows are generated from it. The model therefore doesn’t learn facts about customers, but rather the grammar of relationships.

Something unusual happens when the model is applied. You hand the model the existing table, complete with known answers, along with the new row, all in one go. The model computes through once and delivers the prediction. It does not change its own internal values in the process. Experts call this in-context learning: learning solely from what is currently in the input.

This design has clear limits. Because the entire table must fit into the input, these models work primarily with small datasets, often up to a few tens of thousands of rows. With millions of rows, decision tree methods remain at an advantage. A common misconception is also that a language model is reading tables here. Tabular foundation models are their own separate systems and work with columns of numbers, not with sentences.

From credit checks to the laboratory

In practice, these models turn up wherever small amounts of data carry great value. In medicine, there are often only a few hundred patients per study, and it is precisely there that they demonstrate their strength. The same applies to materials research, chemistry, or quality control in manufacturing. Banks and insurers are also testing them for risk assessments.

In the news, you’ll usually encounter this term under its English name, Tabular Foundation Model. The best-known example is TabPFN from Freiburg, whose second version was presented in the journal Nature in 2025. Major providers such as Google are also working on their own variants. When analysts write that AI is now also conquering the spreadsheet, this technology is essentially what is meant.

Little of this is visible to you directly, since these models are embedded in analysis tools rather than in chat windows. What you should remember is the distinction: a language model writes text, a tabular foundation model fills in missing columns in data tables. Both are pre-trained, but they solve completely different tasks.