Loading page…
The model uses up to 10,000 past examples at prediction time. NVIDIA reports wins over supervised models, while acknowledging that large training sets can favor specialized alternatives.
Listen to this story
NVIDIA made Kumo Relational available on October 21, positioning it as a way to predict from connected database records without building a separate task-specific model. Developers describe prediction targets in PQL, and the system assembles linked historical context rather than requiring a flattened table. NVIDIA reports benchmark gains across 41 tasks and 15 databases, but its in-context setup uses at most 10,000 examples with known outcomes; specialist models can still lead on very large training sets. Deployments include NVIDIA NIM, while code and weights carry different licenses.
On SAP SALT, NVIDIA reports an 8% mean reciprocal rank advantage over AutoGluon ensembles, rising to 16% after fine-tuning.
Evaluations covered classification and regression across RelBench V1 and V2, SAP SALT, and 4DBInfer; NVIDIA says the datasets were absent from pretraining.
The model performed well with 0.2% of training data as context and remained stable after up to 75% of table links were removed.
NVIDIA released Kumo Relational on October 21, promising predictions from linked business tables without training a separate model for each task. The shortcut still depends on history: it reads up to 10,000 examples with known outcomes at prediction time. NVIDIA reports benchmark wins over supervised models, but acknowledges that specialized models can retain an advantage on very large training sets.
The model targets questions such as which customers will leave, how much demand to expect, or which activity may be fraudulent. Rather than requiring developers to flatten a database into one table, it works across connected records—for example, users linked to orders and products—while retaining their timestamps and relationships.
That avoids manually reducing a customer’s transaction history to summaries such as average order value over the past 30 days. NVIDIA argues that those aggregations take substantial work and discard detailed signals in the connections between records. Kumo Relational instead gathers relevant records around each example and produces predictions in a single pass through the model.
Tasks are specified through Predictive Query Language, or PQL, which describes what to predict rather than the steps for calculating it. For time-based tasks, the system reconstructs historical database states. It gives the model records from before a chosen timestamp, then uses later events to establish the example’s outcome, keeping future information out of the predictive inputs.
Inside the model, one network identifies useful information within each table. Another follows the key-based links connecting rows across tables and compares examples with known outcomes. This second stage works on numerical representations of rows rather than every individual cell, a design NVIDIA says supports larger contexts and deeper chains of related records.
NVIDIA evaluated classification and regression tasks across RelBench V1 and V2, SAP SALT, and 4DBInfer.
The evaluation spans e-commerce, medical, enterprise, social and academic data. NVIDIA says none of these datasets appeared in pretraining.
NVIDIA says Kumo Relational beat supervised relational deep-learning models and expert-engineered table-based ensembles across its evaluations. Those comparisons cover both classification, which predicts a category, and regression, which predicts a numerical value. On the SAP SALT enterprise benchmark, it reports an 8% advantage over AutoGluon ensembles in mean reciprocal rank.
Fine-tuning—additional training for a particular dataset and task—increased that reported SAP SALT lead to 16%. NVIDIA also reports resilience when examples or connections are scarce: the model performed well using just 0.2% of training data as context and remained stable with up to 75% of links between tables removed.
The company identifies two important boundaries. Task-specific supervised models may still perform better when training sets are very large; its in-context benchmarks used at most 10,000 examples. The model also needs useful signals to be reachable within a limited number of links. NVIDIA recommends starting with a focused query across a few relevant tables, then expanding as needed.
Kumo Relational is available through NVIDIA’s structured-data-models library and as an NVIDIA NIM deployment. NVIDIA says it can operate in memory, through database and warehouse connectors, or through a custom graph engine handling databases with more than 500 billion rows.
The release separates the code and weight licenses: the code uses Apache 2.0, while the model weights use the OpenMDW License Agreement, version 1.1.
Loading discussion...
Join the conversation
Explain which kinds of decisions would make that shortcut worthwhile.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.