NVIDIA Kumo Relational Predicts From Databases, No Training

Most of the world's business data does not live in documents or images. It lives in relational databases: customers linked to orders, orders linked to products, all changing over time. On October 21, 2026, NVIDIA introduced Kumo Relational, a pre-trained relational foundation model that makes predictions directly on those databases with no task-specific training, no feature engineering and no flattening of tables into one giant spreadsheet. The weights are on Hugging Face and the code is open source.

• What it does: predicts outcomes like churn, demand or fraud in a single forward pass, using up to 10,000 historical examples as context
• Evaluation: 41 tasks across 15 relational databases, none of them seen during pre-training
• Scale: a graph engine handles databases with 500B+ rows via memory-mapped I/O at up to 5 GB/s
• Licensing: code under Apache 2.0; model weights under the OpenMDW License Agreement 1.1

What Is a Relational Foundation Model?

Large language models replaced a zoo of task-specific NLP pipelines with one general model you can simply prompt. NVIDIA argues the same shift is now coming to structured data. Today, building a churn predictor usually means weeks of feature engineering, hand-crafting aggregates like "average order value in the last 30 days," then training a gradient-boosted tree model such as XGBoost for that one task.

Kumo Relational skips those steps. It reads the database as it is, respecting tables, foreign keys and timestamps. You describe the prediction in NVIDIA's Predictive Query Language (PQL), for example counting each user's orders over the next 30 days, and the system samples the relevant subgraph around each entity, gathers labeled historical examples and predicts through in-context learning. It is the structured-data cousin of few-shot prompting.

How Does Kumo Relational Work?

According to NVIDIA's technical write-up, the architecture has three stages:

• Automatic context generation: for time-based tasks it replays historical database states, serving the model only data from before an anchor timestamp to prevent leakage
• Reasoning within tables: a lightweight network alternates column attention and row attention to decide which values matter
• Reasoning across tables: a larger network uses graph attention over primary-key and foreign-key links, interleaved with attention across the context examples

Because the cross-table stage works on row embeddings rather than every cell, NVIDIA says the design avoids quadratic cost and scales to large contexts. The model was pre-trained entirely on synthetic relational databases, generated with an approach based on the PluRel research, so no private customer data was needed.

How Well Does It Perform?

These figures come from NVIDIA's own evaluation, so treat them as vendor-reported until independent groups reproduce them. On RelBench V1 classification tasks, NVIDIA reports Kumo Relational beats gradient-boosted trees on expert-engineered features by nearly 4 AUROC points on average. On the SAP SALT enterprise benchmark it reports a 7-point MRR gain over XGBoost out of the box and 13 points after fine-tuning, plus an 8% MRR lead over AutoGluon ensembles. On the RelArena leaderboard, NVIDIA says it reaches similar accuracy to the top entry while running roughly 430 times faster.

NVIDIA is also candid about limits. The model shines on data with meaningful relationships and time dynamics, such as user behavior. On tasks with very large training sets, task-specific supervised models may still edge it out, which is where optional fine-tuning comes in. Predictive signals also need to sit within a few hops in the graph.

How Can Developers Try It?

There are two routes. Data scientists can install NVIDIA's open-source structured-data-models library from PyPI, which downloads the weights from the Hugging Face Hub and runs on a GPU-backed notebook. Teams that want a service can deploy Kumo Relational as an NVIDIA NIM and query it in PQL. Check the license terms for your specific deployment before going to production.

For context on how fast tabular AI is moving, see our earlier look at SAP's tabular foundation model push, and follow more model releases in our AI coverage.

Sources: Hugging Face Blog: NVIDIA Kumo Relational — October 21, 2026; NVIDIA Build: Kumo Relational model card — October 2026.