NVIDIA Kumo Relational Predicts From Databases

Most of the data companies care about sits in relational databases: customers linked to orders, orders linked to products. Making predictions from it usually means weeks of feature engineering and a separate model for each question. NVIDIA says its new open model can skip that. Kumo Relational, released on 21 October, makes predictions “directly on enterprise databases”, the company says in its launch post on Hugging Face.

Given a question, the model gathers “up to 10,000 historical examples as context and predicts outcomes such as churn, demand, or fraud in a single forward pass”, NVIDIA says. It needs “no task-specific training, no feature engineering, and no flattening of tables”.

The claims, all NVIDIA’s

NVIDIA calls Kumo Relational “the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks”. Across 41 tasks on 15 databases, it says, the model beat both supervised deep-learning models built for relational data and expert-engineered gradient-boosted tree pipelines, on datasets it had not seen during pre-training.

The company gives specific margins. On SAP’s SALT enterprise benchmark, it says the model beats AutoGluon ensembles by 8% in mean reciprocal rank, a ranking measure, rising to 16% after fine-tuning. Against XGBoost on engineered features, a common production set-up, it reports a lead of 7 MRR points on the same benchmark out of the box. On the RelArena-alpha leaderboard, NVIDIA says the model is “about 430x faster than the highest-scoring entry at similar accuracy”, taking roughly 17 seconds per 1,000 predictions against about 7,300.

These are NVIDIA’s own results. The company links to a research paper alongside the release; TSN has not seen independent tests.

How it works

Rather than squashing tables into one flat sheet, the model reads the database as it is, with its tables, links between records and timestamps. A small network first looks within each table; a larger one then reasons across the links, for example tying a user’s profile to their purchase history. Questions are written in what NVIDIA calls Predictive Query Language, which states what to predict rather than how.

The model was pre-trained entirely on synthetic databases, NVIDIA says, generated with an approach called PluRel that members of the Kumo team helped develop. The company says it holds up with as little as 0.2% of the training data as context, and stays stable when up to 75% of the links between tables are removed. It says the model can handle databases of more than 500 billion rows.

Where it falls short

NVIDIA lists limits too. On tasks with very large training sets, it says, task-specific supervised models “may still edge out” its in-context predictions, which used at most 10,000 examples in its tests; fine-tuning is meant for those cases. Useful signals also need to sit within a few links of the record being predicted.

The code is released under the Apache 2.0 licence and the weights under the OpenMDW licence. Developers can run it through NVIDIA’s open-source structured-data-models library or deploy it as an NVIDIA NIM service. The model is also built to be called by AI agents, NVIDIA says, so they can ask which customer is likely to leave and act on the answer.

If the benchmarks hold up outside NVIDIA’s own tests, the bigger change is practical: a prediction becomes a query rather than a project.