Nvidia releases Kumo Relational, a foundation model for database predictions

Nvidia has released Kumo Relational, a pre-trained foundation model designed to make predictions directly on enterprise relational databases. Announced in October 2026, this open-weight model eliminates the need for task-specific training or manual feature engineering by leveraging in-context learning. It represents a significant shift in how machine learning engineers approach structured data, moving away from flattened tables toward native relational reasoning.

What happened

Traditional machine learning pipelines for relational data require extensive preparation. Engineers must flatten complex database schemas into single tables, engineer features manually, and train separate models for each predictive task. This process is labor-intensive and often results in the loss of structural signals present in the original database relationships. Kumo Relational addresses these bottlenecks by operating directly on the database structure, respecting tables, foreign keys, and temporal dynamics without requiring data flattening.

The model was evaluated across 41 tasks on 15 different databases, including benchmarks from RelBench V1 and V2, SAP SALT, and 4DBInfer. These datasets cover diverse domains such as e-commerce, healthcare, and enterprise resource planning. In these tests, Kumo Relational outperformed supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. Notably, on the SAP SALT enterprise benchmark, it surpassed AutoGluon ensembles by 8% in mean reciprocal rank without any fine-tuning.

This release includes both the model weights and the underlying code under open licenses. The weights are available via the Hugging Face Hub under the OpenMDW License Agreement, version 1.1, while the code is released under the Apache 2.0 license. Developers can access the model through Nvidia’s GPU-native structured-data-models library or deploy it as an Nvidia NIM service, allowing for flexible integration into existing infrastructure.

How it works

Kumo Relational functions as a few-shot foundation model that uses in-context learning to adapt to new tasks. Instead of updating model weights through training, it gathers up to 10,000 historical examples as context during inference. Each example consists of a local subgraph around a specific entity at a point in time, paired with its ground-truth target. This allows the model to understand the specific prediction task dynamically based on the provided context.

The architecture processes data in three stages. First, automatic context generation samples entity-centric subgraphs up to an anchor timestamp, ensuring no data leakage by only using historical data. Second, a lightweight network performs reasoning within individual tables, using column and row attention to identify relevant features and filter noise. Finally, a larger network handles reasoning across tables via graph attention over primary and foreign key relationships, linking entities like users to their transaction history efficiently.

To facilitate interaction, the system uses Predictive Query Language (PQL), a declarative interface where users specify what to predict rather than how to compute it. This design avoids the quadratic cost associated with processing every cell in every table by working on row embeddings. As a result, the model scales to databases with over 500 billion rows and maintains high performance even when significant portions of foreign-key links are missing or noisy columns are present.

Key details

• No training required: The model uses in-context learning with up to 10,000 examples, eliminating the need for task-specific training cycles.
• Superior performance: It outperforms supervised models and expert-engineered XGBoost pipelines on 41 benchmarks across 15 databases.
• Native relational handling: Operates directly on database schemas with foreign keys and timestamps, avoiding manual data flattening.
• Scalability: Supports databases with more than 500 billion rows through memory-mapped I/O and efficient graph attention mechanisms.
• Open access: Model weights and code are publicly available, with weights under OpenMDW License v1.1 and code under Apache 2.0.
• Agent ready: Designed to be called by AI agents via PQL, enabling automated predictive workflows alongside generative AI tasks.

Why it matters

For software teams building data-intensive products, this development reduces the barrier to entry for advanced predictive analytics. The traditional workflow of feature engineering and model training often takes weeks and requires specialized data science expertise. By replacing this with a query-based interface, organizations can deploy predictive capabilities for churn, demand forecasting, or fraud detection much faster. This agility allows engineering teams to tackle use cases that were previously too costly or complex to justify.

Furthermore, the ability of foundation models to reason over structured data bridges the gap between large language models and enterprise databases. While LLMs excel at processing unstructured text, they struggle with precise numerical reasoning over complex relational schemas. Kumo Relational provides the predictive layer that allows AI agents to not just read data, but anticipate future states. This integration enables more sophisticated autonomous agents that can act on predicted outcomes, such as proactively addressing customer churn before it happens.

What you can do

• Install the open-source structured-data-models library from PyPI to load the model in a GPU-backed environment.
• Review the quickstart notebook on GitHub to understand how to pass context examples and prediction entities to the model.
• Experiment with Predictive Query Language to define your own predictive tasks without writing custom training pipelines.
• Deploy the model as an Nvidia NIM service if you prefer a separate microservice architecture over in-memory execution.
• Evaluate the model on your own relational datasets using the provided client code for local or Databricks environments.
• Read the associated research paper on arXiv to understand the synthetic pre-training methodology and architectural details.