NVIDIA introduced Kumo Relational, a model that predicts from multi-table databases without task-specific training, reporting results across 41 benchmarks and 15 databases.
NVIDIA introduced Kumo Relational on October 21, 2026, offering developers a model designed to make predictions directly from relational databases. It targets tasks such as estimating customer churn, product demand, or fraud risk, and is available through NVIDIA’s GPU-native open-source library and as an NVIDIA NIM service.
Unlike approaches that first flatten connected tables into a single dataset, Kumo Relational uses database tables, keys, and timestamps as part of its prediction process. NVIDIA says the model is pretrained on synthetic relational data, then adapts to new prediction tasks by examining examples at inference time. Users specify the outcome they want through Predictive Query Language, rather than building a separate task-specific training pipeline.
NVIDIA reports evaluations across 41 tasks and 15 databases spanning areas including e-commerce, healthcare, enterprise, social networks, and academia. The company says those datasets were not used during pretraining, and that the model outperformed the strongest supervised systems tested in classification and regression. On the SAP SALT enterprise benchmark, it exceeded AutoGluon ensembles by 8% in mean reciprocal rank; fine-tuning increased that reported lead to 16%. NVIDIA also reports that the model can work with as little as 0.2% of the training data as context and remain stable when up to 75% of foreign-key links are removed.
The release includes code under the Apache 2.0 license, while model weights use the OpenMDW License Agreement, version 1.1. NVIDIA says its Kumo Engine can handle databases with more than 500 billion rows. The company also notes limits: task-specific supervised models may still perform better with very large training sets, and useful predictive signals must be reachable within a limited number of database connections.