Kumo Relational

NVIDIA · released Oct 21, 2026

UnknownUSstructured relational data

Openness taxonomy

access, licence, and disclosure — never a single open/closed binary

Weights availability: unknown
Licence: unknown
Licence category: unknown
Commercial use allowed: unknown
Derivatives allowed: unknown
Redistribution allowed: unknown
Field-of-use restrictions: unknown
Training code availability: unknown
Inference code availability: unknown
Architecture disclosure: Described as a pre-trained relational foundation model using in-context learning over relational tables, foreign keys, timestamps, and local subgraphs.
Training data disclosure: Pre-trained on synthetic relational data that reflects real-world datasets.

Deployability

raw values as extracted — free text, never inferred from branding

Self-hosting: unknown
On-premise: unknown
Edge deployability: unknown
API availability: Available as an NVIDIA NIM; the text also describes use through NVIDIA's GPU-native structured-data-models library.
Regions available: unknown

Benchmarks

as reported — vendor-supplied unless otherwise noted

41 tasks across 15 databases NVIDIA reports outperforming supervised relational deep learning models and expert-engineered gradient-boosted tree pipelinesSAP SALT enterprise benchmark NVIDIA reports surpassing AutoGluon ensembles by 8% in mean reciprocal rank (MRR)RelBench V1 classification tasks NVIDIA reports outperforming gradient-boosted trees on expert-engineered features by nearly 4 AUROC points on average

National-security pathway & evidence

from the assessed source item, when one is linked

Impact: lowConfidence: mediumNat-sec: noneprimary

Evidence limitations: The source is NVIDIA's announcement and provides no independent reproduction of its benchmark claims or explicit model-weight or licence details.

NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

Linked research papers

from paper_model_links

No linked research papers yet.