NVIDIA Kumo Relational predicts from databases using up to 10,000 in-context examples, no training

NVIDIA has introduced Kumo Relational, a pre-trained relational foundation model that makes predictions directly on enterprise databases from up to 10,000 in-context examples, with no task-specific training, feature engineering, or table flattening. According to NVIDIA, it is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks.

Key facts

• Input: up to 10,000 context examples at inference time. Each pairs a local subgraph around an entity at a point in time with its ground-truth target.
• Inference: one forward pass per prediction.
• Scale: NVIDIA says it scales to databases with 500B+ rows.
• Training data: pre-trained entirely on synthetic relational data, in stages from single tables to complex relational structures.
• Access: NVIDIA's open-source, GPU-native structured-data-models library, or as an NVIDIA NIM.
• Query interface: Predictive Query Language (PQL), a declarative language where the user states what to predict.
• Pricing, parameter count, and release version: not yet disclosed.
• Training cutoff: not applicable in the usual sense, since the model is pre-trained on synthetic data.

Benchmark claims

All figures below come from NVIDIA and its accompanying paper (arXiv 2604.12596). They have not been independently verified.

• Across 41 tasks on 15 databases, NVIDIA claims Kumo Relational beats both supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines.
• On RelBench V1 classification tasks, it outperforms gradient-boosted trees on expert-engineered features by nearly 4 AUROC points on average.
• On the SAP SALT enterprise benchmark, it beats XGBoost on engineered features by 7 MRR points out of the box and by 13 MRR points after fine-tuning.
• On SAP SALT, it surpasses AutoGluon ensembles by 8% in mean reciprocal rank.

Absolute scores were not included in the source material.

Architecture

Most structured-data foundation models operate on a single flat table. Kumo Relational works on the relational schema itself, respecting tables, foreign keys, and timestamps.

1. Automatic context generation. For temporal tasks, the system replays historical database states. It samples entity-centric subgraphs up to an anchor timestamp, computes targets from later events, and gives the model only pre-anchor data. NVIDIA says this prevents data leakage.
2. Reasoning within tables. A lightweight network alternates column attention and row attention on each table independently.
3. Reasoning across tables. A larger network passes row embeddings through graph attention over primary-key/foreign-key relationships, interleaved with cross-sample attention across context examples. Because this stage operates on row embeddings rather than every cell, NVIDIA says it avoids quadratic cost.

Synthetic pre-training

The pre-training generator builds on PluRel (Kothapalli et al., ICML 2026), which Kumo team members co-authored. PluRel generates a schema, foreign-key links between rows that cluster as in real data, and column values from Structural Causal Models with temporal trends and cross-table dependencies. The PluRel paper reports that pre-training loss follows a power law as synthetic databases are added, with gains that carry over to real databases.

What this means

The headline claim is that a single frozen model can replace per-task XGBoost or relational deep learning pipelines on relational data. If the results hold on independent evaluation, that changes the cost of adding a new prediction task from a modeling project to a query. The 13-point MRR gain after fine-tuning on SAP SALT also suggests a path for teams that want more accuracy and can afford task-specific tuning.

The evidence so far is NVIDIA's own paper and benchmarks. Reported gains are relative deltas, and the source gives no latency, hardware requirements, or cost per prediction at 10,000-example context. Those numbers will determine whether the model is practical for high-volume production scoring. The 500B+ row scalability claim also needs third-party validation.

The PQL interface and NIM packaging are aimed at agents. NVIDIA is positioning predictive models as tools agents can call alongside LLMs, which makes tabular prediction a new tool category for agent frameworks.