NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required
huggingface ·
NVIDIA has introduced Kumo Relational, a pre-trained foundation model designed to make predictions directly on enterprise relational databases without requiring task-specific training or feature engineering. By utilizing in-context learning with up to 10,000 historical examples, the model performs single forward-pass predictions for outcomes such as churn, demand, or fraud while respecting database structures like foreign keys and timestamps. Benchmarks across 41 tasks show that Kumo Relational outperforms supervised machine learning models, including expert-engineered gradient-boosted trees and AutoGluon ensembles, achieving higher accuracy with significantly faster inference times. The system operates through a Predictive Query Language interface, allowing both humans and AI agents to query data naturally while avoiding data leakage via temporal replay mechanisms. Built on synthetic relational data generated by the PluRel framework, the model scales to databases containing over 500 billion rows using GPU-native libraries. Although it excels in most scenarios, the authors note that fine-tuning may be necessary for tasks with extremely large training sets where specific supervised models might edge out in-context predictions. The technology is available via NVIDIA's open-source structured-data-models library and can be deployed as an NVIDIA NIM service.
huggingface