- TL;DR
- The Shift to Structured Data Foundation Models
- How Kumo Relational Works
- How Kumo Relational Was Built
- Performance
- Limitations
- Demo
- Start Building with NVIDIA Kumo Relational
- Acknowledgements
TL;DR
NVIDIA Kumo Relational is a pretrained relational foundation model that performs predictions directly on enterprise databases. Given a predictive query, it gathers up to 10,000 historical examples as context and predicts outcomes like churn, demand shifts, or fraud in a single forward pass. It requires no task-specific training, no feature engineering, and no flattening of tables. Kumo Relational is the first few-shot foundation model to outperform traditional supervised machine learning on standard relational benchmarks. Across 41 tasks spanning 15 databases, it surpasses supervised relational deep learning models and expert-tuned gradient-boosted tree pipelines; on the SAP SALT enterprise benchmark, its Mean Reciprocal Rank (MRR) outperforms an AutoGluon ensemble by 8%. It scales to databases with over 500 billion (500B+) rows and is accessible to both humans and agents, available via NVIDIA’s open-source, GPU-native structured-data-models library and deployable as an NVIDIA NIM microservice.
- Paper: https://arxiv.org/abs/2604.12596
- Documentation: https://docs.nvidia.com/sdgm/rfm/overview