The company says it can make predictions such as churn, demand, or fraud directly from related tables, foreign keys, and timestamps without task-specific training, manual feature engineering, or flattening tables.
For each query, Kumo Relational uses up to 10,000 historical examples as in-context data. NVIDIA says the model is available through its newly released open-source, GPU-native structured-data-models library, whose weights download from the Hugging Face Hub, and as an NVIDIA NIM. The company also describes a custom graph engine for databases with more than 500 billion rows.
NVIDIA evaluated the model on 41 tasks across 15 databases from RelBench V1 and V2, SAP SALT, and 4DBInfer. It reports an 8% mean reciprocal rank lead over AutoGluon ensembles on SAP SALT and nearly four average AUROC points over gradient-boosted trees with expert-engineered features on RelBench V1. The model was pre-trained on synthetic relational data. These are NVIDIA's reported results; the source does not establish independent validation, production performance for a specific database, or that it replaces supervised models on every task.