3AM Marketer
AI

NVIDIA Kumo Relational: 430x Faster Inference

NVIDIA Kumo Relational foundation models execute predictive queries directly across multi-table schemas, outperforming gradient-boosted trees and scaling past 500 billion rows.

430x faster
SAP SALT8% MRR
SAP SALT7 MRR points
Out-of-the-Box Lead, from the entry's own table. 5 sources, checked 23 Oct 2026. Full table

NVIDIA Kumo Relational executes predictions directly across multi-table enterprise databases without task-specific training, table flattening, or manual feature engineering pipelines. Across 41 tasks on 15 databases, the relational foundation model outperforms supervised relational deep learning architectures and expert-engineered gradient-boosted tree pipelines.1

Predicting directly over relational databases without table flattening

Traditional relational data machine learning forces teams to write complex SQL join pipelines to flatten normalized tables into static feature matrices before training. Kumo Relational replaces that manual workflow by executing predictive tasks such as churn, demand, or fraud in a single forward pass given a declarative query. It scales to normalized enterprise databases containing more than 500 billion rows.1

Tasks are specified through the Predictive Query Language (PQL), a declarative interface where engineers define what to predict rather than how to calculate tabular features. The model adapts to new tasks at inference time through in-context learning, pairing local subgraphs centered around an entity at a specific point in time with their ground-truth target values. For production workflows integrated with autonomous systems, this architecture allows autonomous AI agents and human analysts to query live databases interchangeably.1

Dual-level attention and temporal sampling eliminate leakage

Within individual tables, a lightweight network processes records independently using column attention to relate values inside a column and row attention to relate features across a single row. Across tables, a larger network distributes row embeddings across the schema using graph attention over primary-key and foreign-key relationships, followed by cross-sample attention across context examples. The underlying weights are pre-trained entirely on synthetic relational data simulated to mirror real-world schemas, relational constraints, and temporal dependencies.1

To prevent target leakage during temporal predictions, Kumo Relational isolates chronological windows before model evaluation. It samples entity-centric subgraphs strictly up to an anchor timestamp and evaluates target labels from events that take place after that cutoff, serving the network only the data recorded prior to the anchor. When evaluating an inference query, the model can incorporate up to 10,000 historical context examples.12

Benchmark numbers: outperforming GBDTs and AutoGluon

Evaluation results across public relational deep learning benchmarks show consistent accuracy gains over established tabular pipelines. On RelBench V1 classification tasks, Kumo Relational outperforms gradient-boosted decision trees trained on expert-engineered features by nearly 4 AUROC points on average.1

Kumo Relational performance deltas against tabular baselines1
BenchmarkBaseline ModelOut-of-the-Box LeadFine-Tuned Lead
SAP SALTAutoGluon Ensembles8% MRR16% MRR
More on SAP SALT
Fine-Tuned Lead
16% MRR
SAP SALTXGBoost (Engineered Features)7 MRR points13 MRR points
More on SAP SALT
Fine-Tuned Lead
13 MRR points
RelBench V1Gradient-Boosted TreesNearly 4 AUROC pointsNot reported
More on RelBench V1
Fine-Tuned Lead
Not reported

On the enterprise SAP SALT benchmark, Kumo Relational outperforms AutoGluon ensembles by 8% in mean reciprocal rank (MRR) out of the box, expanding that advantage to 16% MRR after fine-tuning. Against XGBoost pipelines using engineered features on the same benchmark, Kumo Relational leads by 7 MRR points zero-shot and 13 MRR points after fine-tuning. On the RelArena-α benchmark, Kumo Relational runs approximately 430 times faster than rtl-plurel while matching its predictive accuracy.1

  • Out-of-the-Box Lead7
  • Fine-Tuned Lead13
Kumo Relational lead over XGBoost on SAP SALT in mean reciprocal rank points (2026). (MRR points)1

Benchmarking standards have expanded with RelBench v2, which added four new databases (SALT, RateBeer, arXiv, and MIMIC-IV) along with 36 new tasks and an Autocomplete task format evaluated across independent classification, regression, and recommendation leaderboards.34

How engineering teams deploy it

Engineering teams can access Kumo Relational through NVIDIA's open-source, GPU-native structured-data-models and kumo-engine libraries, or run it in enterprise environments as an NVIDIA NIM microservice. The runtime supports standard tabular tasks including classification, regression, numeric forecasting over a future time horizon, and link prediction to rank or recommend connected entities.15

Documentation page for the Kumo Relational NIM inference service.
Screenshot of kumo.ai, captured 2026-10-23

Teams managing relational machine learning pipelines should evaluate their SQL feature stores against declarative PQL queries on normalized schemas. If tabular jobs bottleneck on manual data flattening or feature drift across foreign keys, in-context relational foundation models are worth testing against them.1