NVIDIA's Kumo Relational outperforms supervised ML on relational benchmarks with no task-specific training

KEY POINTS

1. Kumo Relational is a pre-trained relational foundation model from NVIDIA that makes predictions directly on relational databases by using up to 10,000 historical examples as in-context input and producing outcomes in a single forward pass, without task-specific training, feature engineering, or table flattening.
2. NVIDIA reports it outperforms supervised relational deep-learning models and expert-engineered gradient-boosted pipelines across 41 tasks on 15 databases, beats AutoGluon by 8% MRR on the SAP SALT benchmark, scales to databases with 500B+ rows, and is available via NVIDIA's open-source GPU-native structured-data-models library or as an NVIDIA NIM; it also exposes a Predictive Query Language (PQL) for agent use.
3. If validated, a pre-trained model that outperforms engineered XGBoost pipelines and deep-learning baselines could simplify production workflows, reduce feature engineering and task-specific training, and enable agents to make predictive decisions directly on databases.