NVIDIA has unveiled Kumo Relational, a pre-trained relational foundation model designed to make predictions directly on enterprise databases. This development marks a significant shift in structured data analysis, as the model eliminates the need for task-specific training, feature engineering, or the complex flattening of relational tables.
Direct Predictions via PQL
Unlike traditional machine learning models that require extensive data preparation, Kumo Relational operates directly on existing database structures, respecting foreign keys and timestamps. Using the Predictive Query Language (PQL), users can define what they want to predict—such as customer churn or fraud—and the model gathers up to 10,000 historical examples as context to deliver results in a single forward pass.
Benchmark Superiority
Kumo Relational is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. Key performance metrics include:
- Outperforming expert-engineered gradient-boosted tree pipelines (XGBoost) by nearly 4 AUROC points on RelBench V1.
- Surpassing AutoGluon ensembles by 8% in Mean Reciprocal Rank (MRR) on the SAP SALT enterprise benchmark, with the lead increasing to 16% after fine-tuning.
- Demonstrating high efficiency by running approximately 430 times faster than the highest-scoring entries on the RelArena-α leaderboard at similar accuracy levels.
Scalability and Architecture
The model's architecture alternates between column and row attention within tables and graph attention across the relational structure. This design avoids quadratic costs and allows the model to scale to databases containing over 500 billion rows. Pre-trained on synthetic data generated via the PluRel method, Kumo Relational learns to correlate and aggregate information across tables without relying on fixed, manual aggregations. It is currently available as an open-source library and as an NVIDIA NIM for service-based deployment.
