NVIDIA's Kumo Relational predicts from databases with no training

NVIDIA published Kumo Relational on 21 October 2026, a pre-trained model that makes predictions directly on relational databases. NVIDIA says you state what to predict, such as customer churn, product demand or fraud, and the model answers in a single forward pass without training, feature engineering or flattening tables into one row.

How it works, per NVIDIA

The model reads up to 10,000 historical examples as context at prediction time, each pairing an entity's surrounding records at a point in time with its known outcome. Tasks are written in a declarative Predictive Query Language (PQL). NVIDIA's example is PREDICT COUNT(orders.*, 0, 30, days) FOR EACH users.user_id. For time-based tasks, the post says the system replays the database as it stood before a cut-off date and only shows the model earlier data, to prevent leakage.

NVIDIA says the model was pre-trained entirely on synthetic relational databases, generated with an approach called PluRel, and that it follows the tables, foreign keys and timestamps rather than a single flat table. The post says it scales to databases with more than 500 billion rows.

What NVIDIA reports

NVIDIA reports tests on 41 tasks across 15 databases from RelBench V1 and V2, SAP SALT and 4DBInfer, none seen in pre-training. It says Kumo Relational beats supervised relational deep-learning models and expert-engineered gradient-boosted trees, by nearly 4 AUROC points on average on RelBench V1 classification tasks. On SAP SALT it reports 7 mean-reciprocal-rank (MRR) points over XGBoost on engineered features out of the box, 13 after fine-tuning, and 8% over AutoGluon ensembles.

On the RelArena-α leaderboard, NVIDIA says the model sits on the accuracy-versus-speed frontier at about 17 seconds per 1,000 samples. The top-scoring entry, rt-plurel, reaches Elo 1,878 against Kumo Relational's 1,842 but takes about 7,300 seconds, roughly 430 times longer.

What the post leaves out

The benchmarks and the comparison are NVIDIA's; the post's authors include NVIDIA staff and it thanks the Kumo team, and I found no independent replication. All tests use public research datasets, not production databases. NVIDIA itself notes the model works best on data with meaningful relationships over time, that signals must sit within a few table hops, and that supervised models may still win on tasks with very large training sets. It gives no price, and GPU requirements are only implied by the example setting device="cuda".

What to do

• Pick one existing churn, demand or fraud model and run Kumo Relational against it on the same holdout period, comparing accuracy and the time your team spends maintaining features.
• Check the licences first: NVIDIA lists the code as Apache 2.0 and the weights under the OpenMDW Licence Agreement 1.1.
• Start with a focused query over a few related tables, as NVIDIA advises, and keep the leakage rule: the model should only see data from before the prediction date.