NVIDIA's Kumo Relational Kills the Feature Pipeline

A pre-trained foundation model predicts churn, fraud, and demand directly from enterprise databases—no training, no feature engineering required.

By The ai-compute Desk

NVIDIA's Kumo Relational is a pre-trained relational foundation model that returns predictions in a single forward pass, no task-specific training required. [1]

The model is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. [1] Tested across 41 tasks on 15 databases, it beats expert-engineered gradient-boosted tree pipelines and, on the SAP SALT enterprise benchmark, surpasses AutoGluon ensembles by 8% in mean reciprocal rank. [1]

The architecture explains the advantage. Most structured-data models flatten everything into a single table, discarding the joins and foreign-key relationships where signal lives. Kumo Relational operates directly on the database, respecting its tables, foreign keys, and timestamps, and pulls up to 10,000 historical examples as in-context examples at inference time. [1] It scales to databases with 500B+ rows via memory-mapped I/O at up to 5 GB/s. [1]

The pattern echoes what happened when large language models absorbed the NLP stack. Task-specific pipelines gave way to a single model that generalised. The same logic is now arriving for the rows and columns that underpin most enterprise decision-making.

Practically, a new prediction becomes a query rather than a project. The model is available via NVIDIA's open-source structured-data-models library—with code released under the Apache 2.0 license and model weights under the OpenMDW License Agreement v1.1—or deployable as an NVIDIA NIM. [1]

When fine-tuned, Kumo Relational extends its lead over AutoGluon to 16%—meaning the floor here is not the ceiling. [1]

A new prediction becomes a query instead of a project, so ML teams can finally take on use cases that traditional ML made too laborious to build at every level of the organization.NVIDIA Kumo team