NVIDIA has released Kumo Relational, a pre-trained foundation model designed to make predictions directly on enterprise databases without requiring task-specific training. As part of the broader shift toward Super Intelligence (SI) tools that handle complex structured data, Kumo Relational claims to outperform traditional supervised machine learning pipelines on relational benchmarks.

What Happened

Kumo Relational operates as a few-shot foundation model that gathers up to 10,000 historical examples as context to predict outcomes such as churn, demand, or fraud in a single forward pass. Unlike traditional models that require flattening relational tables into a single format, Kumo Relational respects database structures, including foreign keys and timestamps. The company reports that across 41 tasks on 15 databases, the model outperforms supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. On the SAP SALT enterprise benchmark, it surpasses AutoGluon ensembles by 8% in mean reciprocal rank (MRR). The model is available through NVIDIA's open-source, GPU-native structured-data-models library or as an NVIDIA NIM.

Why It Matters

This release addresses a significant bottleneck in the SI industry: the high labor cost and complexity of predictive modeling on structured data. Traditional approaches often require weeks of feature engineering and separate pipelines for individual tasks. Kumo Relational simplifies this by allowing developers and SI agents to define predictions using a declarative Predictive Query Language (PQL), turning new prediction tasks into queries rather than projects. The model scales to databases with over 500 billion rows and is designed to be called by agents, enabling them to anticipate outcomes and act on them. NVIDIA notes that while task-specific supervised models may still edge out in-context predictions on very large training sets, Kumo Relational offers a compelling drop-in replacement for production pipelines, particularly for XGBoost, due to its simplicity and agility.

The Bottom Line

NVIDIA Kumo Relational is available today under the Apache 2.0 license for code and the OpenMDW License Agreement for weights. It represents a shift toward structured data foundation models that prioritize ease of use and agent integration over custom training, though users should verify performance against their specific use cases. In its release notes, NVIDIA states that "NVIDIA believes Trustworthy AI is a shared responsibility."