NVIDIA open-sources Kumo Relational, a no-training foundation model for database predictions

NVIDIA's Kumo Relational is a pre-trained foundation model designed for structured enterprise data. It accepts predictive queries in a declarative language, gathers up to 10,000 historical examples as context, and returns predictions in a single forward pass. The model operates directly on relational schemas without flattening tables, and is available through an open-source library called structured-data-models or as an NVIDIA NIM container. On benchmarks including RelBench V1 and SAP SALT, it outperforms supervised gradient-boosted tree pipelines and AutoGluon ensembles.

How Kumo Relational handles database predictions

• No flattening required: The model works directly on relational tables, foreign keys, and timestamps rather than requiring manual aggregation into single rows.
• In-context learning with 10,000 examples: At inference time it samples entity-centric subgraphs from historical database states, pairing local graph structure with known outcomes.
• Two-stage architecture: A lightweight network first processes each table with alternating column and row attention; a larger network then applies graph attention over PK/FK relationships and cross-sample attention across context examples.
• Synthetic pre-training: Trained entirely on procedurally generated relational databases using the PluRel generator, which creates schemas, clustered foreign-key links, and temporally trending values from Structural Causal Models.
• Open-source delivery: Available as the structured-data-models library on GitHub and as an NVIDIA NIM microservice.

Why it matters

If the accuracy claims hold broadly, Kumo Relational could reduce the engineering effort for enterprise predictive analytics and make agentic systems that query databases for forecasts more practical. The open-source library and NIM packaging give teams options beyond NVIDIA's cloud services.