Structured data has been the last domain where foundation models mostly did not apply. Language and vision were rewritten by pretraining; the relational database underneath the business was left to per-task models and weeks of feature engineering. NVIDIA’s Kumo Relational aims to close that gap, and the reported numbers are strong enough to take seriously.
Kumo Relational is a pre-trained foundation model that operates directly on a database rather than a flattened table. You describe what you want to predict, it gathers up to 10,000 historical examples as context, and returns a prediction in a single forward pass. No task-specific training, no feature engineering, no hand-rolled aggregation.
The Benchmark Claims
Across 41 tasks on 15 relational databases — RelBench V1 and V2, SAP SALT and 4DBInfer, none seen during pre-training — the model beats the strongest supervised baselines in classification and regression, including relational deep learning models and expert-engineered tabular ensembles. On RelBench V1 classification it outperforms gradient-boosted trees on hand-built features by nearly 4 AUROC points on average. On SAP SALT it beats XGBoost on engineered features by 7 MRR points out of the box and 13 after fine-tuning, and surpasses large AutoGluon ensembles by 8% MRR — a gap widening to 16% with fine-tuning.
The efficiency result may matter more than the accuracy result. On the RelArena-α leaderboard, Kumo Relational sits on the accuracy-efficiency Pareto front at roughly 17 seconds of median inference per 1,000 samples, at an Elo of 1,842. The highest-scoring entry reaches Elo 1,878 — about 36 points better — but takes some 7,300 seconds. That is roughly 430 times slower for a margin most production pipelines would never notice.
How It Works
Pre-training is entirely synthetic, built on PluRel, a relational data generator from the same team. Generation is lightweight, so pre-training data is effectively unlimited, and the underlying paper reports that pre-training loss follows a power law as synthetic databases are added, with gains carrying over to real data. At inference, a lightweight network reasons within tables via alternating column and row attention, and a larger one distributes row embeddings across the schema using graph attention over foreign-key links — working on row embeddings rather than every cell, which keeps the cost from going quadratic.
Tasks are written in NVIDIA’s Predictive Query Language, where you state what to predict rather than how. For temporal tasks the model replays historical database states and serves only pre-anchor data, which prevents the leakage that quietly breaks most tabular evaluations.
Where It Still Loses
NVIDIA is candid: the approach works best where relationships and temporal dynamics carry signal, and on tasks with very large training sets, task-specific supervised models can still edge out in-context prediction. Predictive signals also have to be reachable within a limited number of hops, so queries should start narrow. It ships through NVIDIA’s open-source GPU-native structured-data-models library or as an NIM, with a custom graph engine handling 500B+ rows.
The strategic signal is bigger than the model: XGBoost pipelines are being treated as the incumbent to displace, not the baseline to improve on.
Related: Kumo Relational on Hugging Face