NVIDIA's Kumo Relational predicts straight from relational databases, no training or table flattening required
Synopsis
NVIDIA released Kumo Relational, a pre-trained relational foundation model that predicts directly on enterprise relational databases without task-specific training, feature engineering, or flattening tables into one wide row. Across 41 tasks on 15 databases it beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines, and on the SAP SALT enterprise benchmark it exceeds AutoGluon ensembles by 8% in mean reciprocal rank. A prediction is stated as a declarative query, and the model answers it in a single forward pass using historical examples as context.
Interpretation
Predictive modeling on relational databases gains a path that skips training: given a predictive query, the pre-trained model outputs outcomes such as churn, demand, or fraud without a task-specific model and without manually aggregating multiple tables into one wide table. Earlier practice either required weeks of feature engineering and task-specific training or collapsed links among users, orders, and products into fixed statistics such as average order value in the last 30 days, losing fine-grained structural signals in that step. On 41 benchmarks across 15 databases from RelBench V1 and V2, SAP SALT, and 4DBInfer, the model outperforms the strongest supervised models in both classification and regression, including relational deep learning models and expert-engineered tabular ensembles; none of these datasets appeared in pre-training.
The model reads the database as a database: it respects tables, foreign keys, and timestamps, first understanding values within each table, then linking records along primary-key/foreign-key relationships, and finally relating labeled historical examples to the entity being predicted. The mechanism is in-context learning, meaning historical examples serve as context at inference time rather than updating weights; each example pairs a local subgraph around an entity at a point in time with its ground-truth target. On the SAP SALT enterprise benchmark it surpasses large AutoGluon ensembles by 8% in mean reciprocal rank, and fine-tuning extends the lead to 16%; on RelArena-α it is about 430 times faster than rtl-plurel at similar accuracy.
Pre-training uses entirely synthetic relational data, with a generator that builds each database in three stages: a schema, foreign-key links between rows, and column values with temporal trends and cross-table dependencies, so pre-training data is effectively unlimited. Diverse real relational databases are rarely public, so synthetic data fills the gap; the generator builds on PluRel's extension of single-table generators to full relational databases. Pre-training loss follows a power law as synthetic databases are added, and the gains carry over to real databases; under hard conditions the model performs well with as little as 0.2% of the training data as context, stays stable when up to 75% of foreign-key links are removed, and barely degrades when noisy columns are added.
Perspective
The demonstrated scope is 41 benchmarks on 15 databases spanning e-commerce, medical, enterprise, social, and academic domains, plus the SAP SALT enterprise benchmark and the RelArena-α leaderboard; the model is offered as an open-source library, Hugging Face weights, and a NIM service, with code under Apache 2.0 and weights under the OpenMDW 1.1 license.
On tasks with very large training sets, task-specific supervised models may still edge out in-context predictions; predictive signals must be reachable within a limited number of hops; and the model performs best on data with meaningful relationships and temporal dynamics.