Analysis
NVIDIA researchers describe Kumo Relational as a pre-trained relational foundation model that makes predictions directly on enterprise databases. For each predictive query, written in Predictive Query Language, it gathers up to 10,000 historical examples as context, and the authors say it needs no task-specific training, no feature engineering and no flattening of tables.
The authors evaluated it on 41 benchmarks across 15 databases from RelBench, SAP SALT and 4DBInfer, and call it the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. On SAP SALT they report an 8% gain in mean reciprocal rank over AutoGluon ensembles. Code is released under Apache 2.0 and the weights under the OpenMDW License Agreement, version 1.1; the model is offered through an open-source library or as an NVIDIA NIM.
Analysis: the appeal is speed to a first answer. A churn or demand forecast that once needed a feature pipeline could be tried as a query against tables a business already keeps. The authors' own limits mark the edges of that trial: on tasks with very large training sets, task-specific supervised models may still edge out in-context predictions (which used at most 10,000 context examples in the benchmarks), a case the authors say fine-tuning is designed for; predictive signals must also be reachable within a limited number of hops. A sensible pilot would run the query beside an existing model on the same holdout period, starting with a few relevant tables, before retiring anything.
Source: NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required · 2026-10-21
Why it matters
Test in-context predictions beside an existing model on the same holdout data before treating 'no training required' as a replacement.