NVIDIA's Kumo Relational Predicts Directly on Enterprise Databases Without Training
NVIDIA says its pre-trained Kumo Relational model beats supervised machine learning on relational benchmarks across 41 tasks and 15 databases, returning churn, demand or fraud predictions in a single forward pass with no task-specific training.
In a post published on Hugging Face's blog, the company said the model was evaluated on 41 tasks across 15 databases, where it beat supervised relational deep learning models and gradient-boosted tree pipelines built on expert-engineered features. On RelBench V1 classification tasks it led gradient-boosted trees on engineered features by nearly 4 AUROC points on average. On the SAP SALT enterprise benchmark it beat XGBoost on engineered features by 7 MRR points out of the box and by 13 MRR points after fine-tuning, and surpassed AutoGluon ensembles by 8% in mean reciprocal rank.
Given a predictive query, Kumo Relational gathers up to 10,000 historical examples as context and returns outcomes such as churn, demand or fraud in a single forward pass. Each context example pairs a local subgraph around an entity at a point in time with its ground-truth target. The model works on tables, foreign keys and timestamps rather than requiring them to be flattened into one row through manual aggregation such as "average order value in the last 30 days," an approach NVIDIA said loses fine-grained structural signals and takes significant human effort.
NVIDIA said the model is pre-trained once on synthetic relational data that reflects real-world datasets and adapts to each new task through in-context learning at inference time. It scales to databases with more than 500 billion rows and is available to both people and agents through the company's open-source, GPU-native structured-data-models library or as an NVIDIA NIM. A Predictive Query Language gives agents an interface for posing predictive questions, and a prompting interface converts natural language into a predictive query.
NVIDIA framed the release as part of a shift already seen in language and vision, where a single pre-trained model displaced per-task pipelines for reasons of accuracy, simpler deployment and flexibility. Applied to structured data, the company argued that a pre-trained model can replace separate pipelines for individual tasks so that a new prediction becomes a query rather than a project, and can give LLM-based agents the ability to anticipate outcomes and act on them. The announcement was accompanied by a paper at arXiv 2604.12596 and technical documentation.