NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
NVIDIA’s Kumo Relational foundation model outperforms expert-engineered XGBoost pipelines by 4 AUROC points and AutoGluon by 8% MRR on relational databases without requiring any task-specific training, feature engineering, or table flattening. For production engineers, this replaces fragmented, high-maintenance tabular ETL pipelines with a single NVIDIA NIM that can run zero-shot predictive queries directly on databases containing over 500 billion rows. This shifts structured data modeling from a weeks-long engineering cycle to a real-time API call, allowing LLM agents to perform zero-shot business predictions natively.
NVIDIA Kumo Relational outperforms supervised ML on relational tasks by 8% MRR on SAP SALT benchmarks and scales to databases with 500B+ rows without requiring task-specific training or feature engineering. This allows engineers to replace complex ML pipelines with a single pre-trained model, enabling faster, more accurate predictions directly on enterprise databases and reducing development time from weeks to mere queries.