NVIDIA's Kumo Relational predicts from databases with no training

NVIDIA has released Kumo Relational, a pre-trained model that makes predictions directly on business databases without task-specific training. The details come from a post on Hugging Face by NVIDIA researchers, published on 21 October ^[1]. The benchmark results are NVIDIA's own, taken from its accompanying paper, and have not been independently verified here.

What it does

Most predictive modelling on company data means flattening linked tables into one row per customer, then engineering features and training a model, which NVIDIA says can take weeks ^[1]. Kumo Relational works on the database itself, respecting its tables, links and timestamps. You describe what to predict, such as customer churn, demand or fraud, in a simple query language, and the model returns predictions in a single pass ^[1].

It learns each new task from examples. For a given question it gathers historical cases from the database, each pairing a customer or item at a point in time with what actually happened, and reasons over them. It was pre-trained once on synthetic relational data rather than real company databases ^[1].

The claims

NVIDIA says that across 41 tasks on 15 databases, none seen in pre-training, the model beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines ^[1]. On the SAP SALT enterprise benchmark it surpasses AutoGluon ensembles by 8% in mean reciprocal rank, extending to 16% after fine-tuning ^[1]. On one set of classification tasks it beats expert-engineered tree models by nearly 4 AUROC points on average ^[1]. NVIDIA also says it performs well with as little as 0.2% of the training data as context and can scale to databases with more than 500 billion rows ^[1].

Limits NVIDIA lists

The post is candid about weaknesses. The model performs best on data with meaningful relationships and time-based behaviour. On tasks with very large training sets, task-specific supervised models may still do slightly better. And predictive signals must be reachable within a limited number of table "hops" ^[1].

Availability

NVIDIA says Kumo Relational is available now. The code is released under the Apache 2.0 licence and the weights under the OpenMDW License Agreement, version 1.1. It can run through NVIDIA's open-source library on a GPU, or as an NVIDIA NIM service ^[1]. NVIDIA also says AI agents can call it to anticipate outcomes, for example which customer will churn ^[1].

What it means for businesses

Many firms hold their most valuable data in linked tables of customers, orders and products, yet lack the data-science capacity to build a model for every question. A single model that treats a new prediction as a query rather than a project could widen who can use predictive analytics.

Caution is sensible. Benchmarks are not your data, and running it needs GPU infrastructure and technical staff. A reasonable first step is a pilot on one familiar question, such as churn, compared against your current method.

Source: Hugging Face (NVIDIA) — NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required