NVIDIA releases Kumo Relational, a no-training foundation model for relational data

NVIDIA researchers published Kumo Relational on the Hugging Face blog, a pre-trained relational foundation model that makes predictions directly on enterprise databases. A user connects a database and describes what to predict in a predictive query language; the model returns outcomes such as churn, demand or fraud in a single forward pass, with no task-specific training, no feature engineering and no flattening of tables.

The release reflects NVIDIA's bet that structured data is the next domain to be rewritten by foundation models. Relational databases still demand weeks of feature engineering and one model per task. NVIDIA describes Kumo Relational as the first few-shot foundation model to beat supervised machine learning on common relational benchmarks, and says it scales to databases with more than 500 billion rows.

At inference the model reads up to 10,000 historical examples as context, each pairing a local subgraph around an entity at a point in time with its ground-truth outcome. A lightweight network reasons within each table using alternating column and row attention, then a larger network propagates row embeddings across primary-key and foreign-key links with graph attention, interleaved with cross-sample attention. Because that stage works on row embeddings rather than on every cell, cost does not grow quadratically with table size.

Pre-training uses entirely synthetic relational data. A generator built on the PluRel approach creates schemas, foreign-key links and column values from structural causal models with temporal trends, and pre-training moves from single tables to increasingly complex structures. NVIDIA says the synthetic route makes pre-training data effectively unlimited, and that PluRel showed pre-training loss follows a power law as databases are added.

The evaluation spans 41 tasks on 15 databases from RelBench V1 and V2, SAP SALT and 4DBInfer, none of them seen in pre-training. On RelBench V1 classification tasks it beats gradient-boosted trees on expert-engineered 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 AutoGluon ensembles by 8% in mean reciprocal rank.

Efficiency is part of the pitch. On the RelArena-alpha leaderboard it reaches an Elo of 1,842 at about 17 seconds per 1,000 samples, while the top entry, rt-plurel, scores 1,878 but takes roughly 7,300 seconds - about 430 times slower. The team also reports that performance holds with as little as 0.2% of training data as context, with up to 75% of foreign-key links removed, and with many noisy columns added.

The blog is candid about limits. With very large training sets, task-specific supervised models may still edge out in-context predictions, which is where fine-tuning comes in. Predictive signals also have to be reachable within a limited number of hops, so NVIDIA recommends starting with a focused query over a few relevant tables and expanding the graph from there.

Kumo Relational is available through NVIDIA's open-source, GPU-native structured-data-models library or as an NVIDIA NIM. Its Predictive Query Language is designed for humans and agents alike, which means an agent can treat a question about next month's churn as a query rather than a project. The open question is whether the benchmark lead survives messy production databases, and whether teams stop treating feature engineering as the default first step.