The Pulse
NVIDIA Says Kumo Relational Beat Supervised Models on 41 Tasks
NVIDIA says its Kumo Relational model outperformed supervised machine-learning systems across 41 benchmarks on 15 databases. The model predicts from connected database tables using historical examples as context, without task-specific train
AI.info Team · Written with AI · fact-checked ·
NVIDIA says its Kumo Relational model outperforms supervised machine-learning systems on 41 prediction tasks across 15 databases, without training a separate model for each task. The company announced the release on October 21, 2026, describing it as a foundation model for predictions from connected database tables. The results come from benchmarks spanning e-commerce, medicine, enterprise, social networks and academic data—not a claim that the model wins on every production database.
Kumo Relational reads connected tables as a graph
Many models for structured data expect one flat table. Business records, however, often sit across linked tables: users connect to orders, and orders to products. Flattening those links into a single row can require manual summaries, such as average order value over a set period, while discarding details about individual records.
Kumo Relational instead uses table rows, columns, foreign-key connections and timestamps. For each prediction, it draws on up to 10,000 historical examples as context, then produces an output in one forward pass. NVIDIA says developers can ask for tasks such as predicting customer churn, demand or fraud without task-specific training, feature engineering or flattening the tables.
The underlying research paper, KumoRFM-2: Scaling Foundation Models for Relational Learning, describes the model family behind the release. Its authors write:
“Finally, while KumoRFM-1 was limited to small-scale in-memory datasets, KumoRFM-2 scales to billion-scale relational datasets.”
Valter Hudovernik et al., KumoRFM-2
What NVIDIA’s benchmarks measure
NVIDIA evaluated the model on 41 tasks across 15 databases from RelBench versions 1 and 2, SAP SALT and 4DBInfer. The company says those datasets were not included in pretraining, and that Kumo Relational outperformed the strongest supervised models it tested, including relational deep-learning systems and ensembles built on engineered tabular features.
On the SAP SALT enterprise benchmark, NVIDIA reports an 8% advantage in mean reciprocal rank over large AutoGluon ensembles, rising to 16% after fine-tuning. The company also says the model outperformed engineered-feature XGBoost systems by seven mean reciprocal rank points out of the box and 13 points after fine-tuning. Those are benchmark results reported by NVIDIA; they do not establish that the model will beat a carefully tuned system on any particular company’s data.
NVIDIA’s RelArena-α comparison puts Kumo Relational on an accuracy-and-speed trade-off frontier. The company says it reached similar accuracy to the leaderboard’s highest-scoring entry while running about 430 times faster in that comparison. That result describes the tested workloads and systems, not a general guarantee about inference speed in deployment.
Data size and task fit still matter
NVIDIA says the model works with a custom graph engine that can handle databases with more than 500 billion rows using memory-mapped input at up to 5 GB per second. That figure describes the engine’s stated capacity, not a benchmark showing that each prediction processes 500 billion rows. The model is also available through NVIDIA’s GPU-native structured-data-models library and as an NVIDIA NIM service.
The announcement names limits as well as gains. Kumo Relational performs best when the data contains useful relationships and changes over time; signals also need to be reachable within a limited number of table-link steps. NVIDIA says specialized supervised models may still do better on tasks with very large training sets, where the model’s benchmark context was capped at 10,000 examples.
The code is released under Apache 2.0, while the model weights use the OpenMDW License Agreement, version 1.1, according to NVIDIA. The practical test for teams will be whether their own linked data, prediction targets and deployment setup match the conditions where the benchmark results were strongest.