NVIDIA's Kumo Relational Predicts from Live Databases Without Task-Specific Training
If Kumo Relational holds up outside NVIDIA's benchmarks, the weeks of feature engineering and pipeline tuning that predictive modeling on relational data has required are the part of the workflow this model is built to remove.
Reporting from 1 source: GIGAZINE.
NVIDIA released Kumo Relational, a pretrained foundation model that makes predictions directly against corporate relational databases. Given a prediction query, it gathers up to 10,000 past cases as context and forecasts churn, supply and demand, and fraud in a single instruction. NVIDIA says it needs no task-specific training, no feature engineering, and no flattening of tables, and that it surpassed supervised machine learning models on benchmarks of common relational databases.
The model is pretrained once on synthetic relational data that mirrors real-world datasets, then adapts to new tasks through in-context learning. At inference it reads tables, foreign keys, and timestamps as they are stored, and NVIDIA says it handles everything from building the context to running the model.
NVIDIA's argument for the design is that real corporate data lives across linked tables, where users connect to purchase histories and purchase histories connect to products, and all of it changes over time. Flattening that into a single row by hand loses detail and costs labor. Kumo Relational is aimed at that aggregation step, not at the query itself.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.