NVIDIA Kumo Relational Outperforms Supervised Machine Learning Without Training

NVIDIA has introduced Kumo Relational, a pre-trained relational foundation model that makes predictions directly on enterprise databases without task-specific training, feature engineering, or table flattening. Given a predictive query, the model retrieves up to 10,000 historical examples as in-context samples and predicts outcomes such as churn, demand, or fraud in one forward pass. It is intended as a structured-data counterpart to language foundation models, aiming to replace labor-intensive pipelines for relational databases. Kumo Relational operates on database schemas, tables, foreign keys, and timestamps, preserving relationships among entities such as users, orders, and products. It uses a Predictive Query Language for declarative tasks and generates context by sampling entity-centered subgraphs from historical database states, computing targets from future events while preventing data leakage. The architecture reasons within tables through column and row attention and across tables with graph attention and cross-sample attention, allowing scalable relational inference without manually fixed aggregations. The model is pre-trained on synthetic relational data generated to reflect real-world databases and temporal dynamics. Pre-training stages progress from single-table to complex relational structures, enabling it to learn filtering, correlation, and aggregation across tables. When higher accuracy is needed, it can be fine-tuned on a specific dataset and task. On 41 benchmarks across 15 databases, including RelBench, SAP SALT, and 4DBInfer, NVIDIA says Kumo Relational outperforms supervised relational deep-learning models and expert-engineered gradient-boosted tree pipelines in classification and regression. It surpasses AutoGluon ensembles by 8% in mean reciprocal rank on SAP SALT and improves to 16% after fine-tuning. NVIDIA positions the model on the accuracy-efficiency Pareto front, claiming it is far faster than the highest-scoring alternative while maintaining competitive accuracy. NVIDIA said the model scales to databases with more than 500 billion rows through a custom graph engine with memory-mapped I/O, and it can support both human analysts and software agents. It is available in an open-source GPU-native structured-data library, as downloadable weights, and as an NVIDIA NIM microservice for separate deployment. The model weights are released under the OpenMDW License Agreement, version 1.1, while code is under Apache 2.0. The company says predictive signals should be reachable within a limited graph radius and recommends starting with focused queries over relevant tables.