What changes compared with traditional machine learning
Nvidia has introduced Kumo Relational, described in the Hugging Face blog as a pretrained foundation model that makes predictions directly on enterprise databases. According to the source, the key point is that it requires no task-specific training, no manual feature engineering, and no reduction of tables to a single flat format. The user formulates a predictive request, and the model gathers up to 10,000 historical examples as context, producing the prediction in a single pass.
The reported benchmark results
The company claims that Kumo Relational is the first few-shot model of its kind to outperform supervised machine learning on common relational benchmarks. Across 41 tasks spread over 15 databases, it reportedly surpasses supervised relational deep learning models and gradient boosting pipelines built with expert-designed features. On the enterprise benchmark SAP SALT, it would outperform AutoGluon ensembles by 8% on the MRR metric. On RelBench V1 classification tasks, the average advantage would be nearly 4 AUROC points over boosted trees with engineered features, and on SAP SALT 7 MRR points without tuning, up to 13 after optimization.
The model is also said to be able to scale to databases with more than 500 billion rows and to be accessible both through an open-source GPU-native library and as an Nvidia NIM service. According to the source, it is also designed to be queried by software agents, thanks to a declarative predictive query language.
Why the challenge is technical and what remains to be verified
The main difficulty, the material explains, lies in the fact that real data live across multiple tables linked by foreign keys and temporal dimensions. Flattening them into a single row with manual aggregations loses structural signals and requires a great deal of human work. The model tackles the problem with separate attention to column values, rows, and relationships between tables, reconstructing the historical state of the database up to a reference point in time to prevent future information from contaminating the prediction.
Training takes place on entirely synthetic relational data, generated to reflect real databases. Because the latter are rarely public, synthetic data fill a gap, but it is precisely this step that merits independent verification. The source is the manufacturer's presentation: the figures cited come from internal evaluations and are not shown in the material to have been replicated by third parties. It therefore remains to be confirmed how well the model generalizes to enterprise contexts other than the benchmarks, along with its operating costs and robustness outside the tested domains.
