NVIDIA Releases Kumo Relational Model

NVIDIA announced Kumo Relational on October 21, 2026, as a pre-trained model that predicts outcomes directly from enterprise relational databases. Given a predictive query, it gathers up to 10,000 historical examples as context and returns predictions such as customer churn, demand, or fraud in a single forward pass. NVIDIA says it requires no task-specific training, feature engineering, or table flattening, and is designed for use by people and software agents.

The model operates on database tables, foreign keys, and timestamps rather than converting related records into one flat table. It is pre-trained on synthetic relational data and adapts to a prediction task through in-context learning (using examples provided with a query), pairing each example’s entity-centered data at a point in time with its known outcome. For temporal tasks, it uses database information from before an anchor timestamp to predict events after it, a design intended to prevent data leakage.

NVIDIA evaluated Kumo Relational on 41 tasks across 15 databases from RelBench V1 and V2, SAP SALT, and 4DBInfer, covering e-commerce, medical, enterprise, social, and academic data not used in pre-training. The company reports that it outperformed the strongest supervised models in classification and regression, including relational deep-learning models and ensembles built with expert-designed features. On RelBench V1, it exceeded gradient-boosted trees by nearly 4 AUROC points on average; on SAP SALT, it beat XGBoost by 7 MRR points without fine-tuning and 13 after fine-tuning.

On the RelArena-α leaderboard, NVIDIA says Kumo Relational reached Elo 1,842 at about 17 seconds per 1,000 samples, compared with Elo 1,878 at about 7,300 seconds for the highest-scoring entry, rt-plurel; the company describes Kumo as roughly 430 times faster at similar accuracy. On SAP SALT, it exceeded large AutoGluon ensembles by 8% in mean reciprocal rank (MRR) and recent tabular foundation models by about 25%; fine-tuning widened its lead over AutoGluon to 16%. It performed with 0.2% of training data as context, remained stable after removal of up to 75% of foreign-key links, and barely degraded when noisy columns were added.

NVIDIA says the model works best when data has meaningful relationships and changes over time. For tasks with very large training sets, task-specific supervised models may still outperform its in-context predictions, which used at most 10,000 examples in the benchmarks; fine-tuning is offered for such cases. Predictive signals must also be reachable within a limited number of relationship hops, so NVIDIA recommends beginning with a focused query over a few tables. The model is available through NVIDIA’s open-source, GPU-native structured-data-models library or as an NVIDIA NIM service. A custom graph engine supports databases with 500B+ rows using memory-mapped input/output at up to 5 GB/s. Code is under the Apache 2.0 license, while model weights use the OpenMDW License Agreement, version 1.1.

NVIDIA announced Kumo Relational on October 21, 2026, as a pre-trained model that predicts outcomes directly from enterprise relational databases. Given a predictive query, it gathers up to 10,000 historical examples as context and returns predictions such as customer churn, demand, or fraud in a single forward pass. NVIDIA says it requires no task-specific training, feature engineering, or table flattening, and is designed for use by people and software agents.

The model operates on database tables, foreign keys, and timestamps rather than converting related records into one flat table. It is pre-trained on synthetic relational data and adapts to a prediction task through in-context learning (using examples provided with a query), pairing each example’s entity-centered data at a point in time with its known outcome. For temporal tasks, it uses database information from before an anchor timestamp to predict events after it, a design intended to prevent data leakage.

NVIDIA evaluated Kumo Relational on 41 tasks across 15 databases from RelBench V1 and V2, SAP SALT, and 4DBInfer, covering e-commerce, medical, enterprise, social, and academic data not used in pre-training. The company reports that it outperformed the strongest supervised models in classification and regression, including relational deep-learning models and ensembles built with expert-designed features. On RelBench V1, it exceeded gradient-boosted trees by nearly 4 AUROC points on average; on SAP SALT, it beat XGBoost by 7 MRR points without fine-tuning and 13 after fine-tuning.

On the RelArena-α leaderboard, NVIDIA says Kumo Relational reached Elo 1,842 at about 17 seconds per 1,000 samples, compared with Elo 1,878 at about 7,300 seconds for the highest-scoring entry, rt-plurel; the company describes Kumo as roughly 430 times faster at similar accuracy. On SAP SALT, it exceeded large AutoGluon ensembles by 8% in mean reciprocal rank (MRR) and recent tabular foundation models by about 25%; fine-tuning widened its lead over AutoGluon to 16%. It performed with 0.2% of training data as context, remained stable after removal of up to 75% of foreign-key links, and barely degraded when noisy columns were added.

NVIDIA says the model works best when data has meaningful relationships and changes over time. For tasks with very large training sets, task-specific supervised models may still outperform its in-context predictions, which used at most 10,000 examples in the benchmarks; fine-tuning is offered for such cases. Predictive signals must also be reachable within a limited number of relationship hops, so NVIDIA recommends beginning with a focused query over a few tables. The model is available through NVIDIA’s open-source, GPU-native structured-data-models library or as an NVIDIA NIM service. A custom graph engine supports databases with 500B+ rows using memory-mapped input/output at up to 5 GB/s. Code is under the Apache 2.0 license, while model weights use the OpenMDW License Agreement, version 1.1.