Glossary

What is Relational foundation model?

A relational foundation model is a pre-trained AI model that makes predictions directly from several linked database tables, such as customers, contracts and claims. Unlike classic machine learning, it does not need the tables to be laboriously flattened into one, and often needs no separate training for each task. A well-known example is KumoRFM.

In simple terms

A language model has read so much text that it understands new text without learning from scratch for each question. In the same way, a relational foundation model has seen many databases and knows typical patterns: how customers, orders and payments relate, and what often comes before a cancellation. Instead of a spreadsheet where someone copies everything into one list by hand, it reads the database as it is.

In the enterprise

Muster Handels GmbH wants to know which business customers are likely to order less over the next three months. The data sits in several tables: customers, orders, invoices, complaints. Instead of spending weeks building features for a custom model, the team connects the tables to a relational foundation model and states the prediction question. The results are checked against historical data before sales uses them.

Related terms

Sources

  1. NVIDIA Kumo documentation: KumoRFM: A Relational Foundation Model
  2. Fey et al. (2026): KumoRFM-2: Scaling Foundation Models for Relational Learning (arXiv:2604.12596)
  3. Robinson et al. (2024): RelBench: A Benchmark for Deep Learning on Relational Databases (arXiv:2407.20060)

How we use this: UNID AI Predict · Predictive and analytical AI on your enterprise data

← All terms

See it working on your own material.

In a 30-minute call we show the products live and work out which use case makes sense for you first. No obligation, in German or English, in Vienna or online.

UNID AI · Leonard-Bernstein-Straße 10 · 1220 Vienna · +43 660 476 30 89