No feature engineering, no flattening, no training
Nvidia has released Kumo Relational, a pre-trained foundation model that makes predictions directly on a relational database. You connect the database, write a predictive query in a language Nvidia calls PQL, and the model gathers up to 10,000 historical examples as context and predicts the outcome in one forward pass.
What it skips is the part that normally takes the time. Predicting churn or fraud from a multi-table database usually means flattening the tables into one wide file, hand-engineering features from the joins, and training a model on them. Kumo Relational does none of that. It was pre-trained once on synthetic relational data generated by a method Nvidia calls PluRel, and each new task runs through in-context learning rather than training. For temporal tasks it replays historical database states so the model cannot see the future.
The results are Nvidia’s own
Nvidia describes it as the first few-shot foundation model to beat supervised machine learning on common relational benchmarks, and the numbers come from the company’s own paper. They have not been independently reproduced.
The evaluation covers 41 tasks across 15 databases drawn from RelBench V1 and V2, SAP SALT and 4DBInfer, spanning e-commerce, medical, enterprise, social and academic data, none of it seen in pre-training. On RelBench V1 classification, Nvidia reports the model running nearly four AUROC points ahead of gradient-boosted trees on expert-engineered features. On SAP SALT it reports an 8% better mean reciprocal rank than large AutoGluon ensembles, and seven MRR points better than XGBoost on engineered features out of the box, widening to 13 after fine-tuning.
The robustness figures are the more interesting ones. Nvidia says the model performs well with as little as 0.2% of the training data as context, holds up when three-quarters of the foreign-key links are removed, and barely degrades when noisy columns are added. Those are the conditions real databases arrive in.
The trade on the leaderboard
On the RelArena-α leaderboard Nvidia places the model at roughly Elo 1,842 at about 17 seconds per 1,000 samples. The top entry, rt-plurel, reaches Elo 1,878 — and takes about 7,300 seconds for the same work, roughly 430 times longer. That is the Pareto argument the release rests on: a small accuracy gap bought back with two and a half orders of magnitude of speed.
Where it stops working
Nvidia names the limits plainly. The model wants data with real relationships and temporal dynamics; on tasks with very large training sets, a supervised model may still win, since in-context prediction is capped at 10,000 examples in these benchmarks. Predictive signal has to be reachable within a limited number of hops across the schema, so the documentation recommends starting with a focused query over a few tables rather than pointing it at everything.
The code is Apache 2.0 and installs as a Python package; the weights sit on Hugging Face under version 1.1 of the OpenMDW License Agreement and download on first use. It can also run as an Nvidia NIM.
What to watch is whether anyone outside Nvidia reproduces the RelBench margins. A foundation model that removes feature engineering from enterprise prediction is a large claim, and so far the only party that has tested it is the party selling the GPUs it runs on.
