NVIDIA says Kumo Relational predicts from enterprise databases without training and beats supervised ML benchmarks

NVIDIA introduced Kumo Relational, a pre-trained relational foundation model that makes predictions directly on enterprise databases from Predictive Query Language queries, using up to 10,000 historical examples as context in a single forward pass. NVIDIA says it needs no task-specific training, feature engineering, or table flattening, and reports that it beat supervised relational deep learning and engineered gradient-boosted tree pipelines across 41 tasks on 15 databases, including an 8% MRR gain over AutoGluon ensembles on SAP SALT. It is offered through NVIDIA's open-source GPU-native structured-data-models library or as an NVIDIA NIM, with claimed scaling to databases above 500B rows.

Teams maintaining XGBoost-style relational pipelines could test Kumo as a drop-in baseline for high-effort churn, demand, or fraud tasks because the reported gains came without per-task feature engineering. The trade-off is operational: predictions depend on building context from up to 10,000 historical examples and running NVIDIA's GPU/NIM stack, while the article does not provide deployment latency, cost, or results on a team's own schema and data quality constraints.