NVIDIA Kumo Relational Beats XGBoost on Enterprise Databases Without Training
NVIDIA's pre-trained relational foundation model outperforms supervised ML and gradient-boosted trees on 41 benchmark tasks, requiring no feature engineering or task-specific training.
Last verified:
A Foundation Model for Tables
NVIDIA released Kumo Relational, a pre-trained foundation model designed to predict outcomes directly from relational databases without training or feature engineering. According to the Hugging Face Blog, the model gathers up to 10,000 historical examples as context and generates predictions for tasks like customer churn, demand forecasting, and fraud detection in a single forward pass. Kumo Relational is the first few-shot foundation model to outperform supervised machine learning on relational benchmarks.
Benchmark Performance Against Production Systems
The model’s results challenge incumbent approaches. Across 41 tasks spanning 15 databases, Kumo Relational beats both relational deep learning models and expert-engineered gradient-boosted tree pipelines. On the SAP SALT enterprise benchmark, it surpasses AutoGluon ensembles by 8% in mean reciprocal rank (MRR). According to Hugging Face, on RelBench V1 classification tasks, Kumo Relational outperforms gradient-boosted trees trained on expert-engineered features by nearly 4 AUROC points on average. The gap widens after fine-tuning: the model beats XGBoost on engineered features by 13 MRR points on SAP SALT, compared to 7 MRR points without additional tuning.
Scaling and Deployment
Kumo Relational scales to databases with 500 billion or more rows, addressing a constraint that has historically limited structured-data models to smaller enterprise systems. The model is available through two channels: NVIDIA’s open-source, GPU-native structured-data-models library (supporting local deployment on GPU infrastructure) and NVIDIA NIM, a managed inference service. According to Hugging Face, the model is designed to work with both human teams and AI agents, offering a Predictive Query Language (PQL) interface for agents to pose predictive questions directly.
Why This Matters
The release signals a shift in how teams approach enterprise predictive modeling. If Kumo Relational’s benchmark performance reproduces in production, organizations relying on XGBoost pipelines—which require weeks of feature engineering and task-specific tuning—face a clear economic incentive to migrate. The elimination of feature engineering reduces ML operations friction and accelerates time-to-prediction, especially valuable for teams supporting multiple use cases. For agentic AI systems, the addition of predictive capability (rather than just read-write-reason) extends the scope of autonomous decision-making in data-driven workflows. The remaining open questions are production latency (query-to-response time for real-time scoring), cold-start performance on novel databases, and whether the SAP SALT results generalize to vertical-specific relational schemas outside the benchmark.
Frequently Asked Questions
Does Kumo Relational require feature engineering or training?
No. It is a pre-trained foundation model that makes predictions directly on enterprise databases without task-specific training, feature engineering, or table flattening.
How does Kumo Relational's performance compare to XGBoost?
According to Hugging Face, Kumo Relational outperforms expert-engineered XGBoost pipelines by 7 mean reciprocal rank (MRR) points on the SAP SALT enterprise benchmark out of the box, and by 13 MRR points after fine-tuning.
What is the maximum database size Kumo Relational can handle?
The model scales to databases with 500 billion or more rows.
How can teams integrate Kumo Relational into production?
NVIDIA offers two deployment paths: the open-source, GPU-native structured-data-models library or NVIDIA NIM (a managed inference service).