NVIDIA Kumo Relational Beats Supervised ML Without Task Training
NVIDIA Kumo Relational predicts directly from databases without task training, beating supervised models across 41 tasks and scaling past 500 billion rows.
Summary
NVIDIA released Kumo Relational on October 21, 2026, calling it the first few shot relational foundation model to outperform supervised machine learning. Pretrained entirely on synthetic databases using an approach based on PluRel from ICML 2026, it processes tables, foreign keys and timestamps directly, without feature engineering, flattening or task specific training. Predictive Query Language queries trigger one forward pass using up to 10,000 historical examples, with column, row, graph and cross sample attention preventing leakage and preserving relational structure.
Across 41 classification and regression tasks on 15 unseen databases from RelBench V1 and V2, SAP SALT and 4DBInfer, Kumo beat supervised relational deep learning and engineered tabular ensembles. It led engineered boosted trees by nearly 4 AUROC points on RelBench V1; on SAP SALT, it beat XGBoost by 7 MRR points out of the box and 13 after fine tuning, AutoGluon by 8%, rising to 16% after fine tuning, and recent tabular foundation models by about 25%. It remained effective with 0.2% of training data and 75% of foreign key links removed. Available through NVIDIA’s Apache 2.0 structured-data-models library, OpenMDW 1.1 weights and NVIDIA NIM, it handles more than 500 billion rows at up to 5 GB/s. Very large training sets can still favor supervised models, while useful signals must lie within limited graph hops.
Positives
- 41 tasks across 15 unseen databases showed Kumo Relational beating supervised relational deep learning and engineered tabular ensembles.
- Nearly 4 AUROC points separated Kumo Relational from engineered boosted trees on average across RelBench V1 classification tasks.
- 8% higher MRR than AutoGluon on SAP SALT increased to a 16% lead after task specific fine tuning.
- 0.2% of training data remained sufficient for useful context, while removing up to 75% of foreign key links caused limited disruption.
- More than 500 billion rows can be handled through a graph engine delivering memory mapped I/O at up to 5 GB/s.
- Apache 2.0 code, OpenMDW 1.1 model weights, NVIDIA NIM deployment and Hugging Face distribution are available now.
Risks & concerns
- 10,000 context examples cap in context inference, allowing task specific supervised models to edge ahead on very large training sets.
- Limited graph hops constrain reachable predictive signals, requiring focused queries over a few relevant tables before expansion.
- Meaningful relationships and temporal dynamics produce the best results, limiting suitability for data lacking those characteristics.