NVIDIA Releases Kumo Relational: Zero-Training Predictions on Enterprise Databases
What changed
NVIDIA has released Kumo Relational, a pre-trained relational foundation model designed to generate predictions directly on enterprise databases without requiring task-specific training, feature engineering, or table flattening. The model accepts predictive queries and utilizes up to 10,000 historical examples as context to produce results in a single forward pass.
Benchmarking across 41 tasks on 15 databases demonstrates significant performance gains over established supervised baselines. On RelBench V1 classification tasks, Kumo Relational exceeds gradient-boosted tree pipelines built on expert-engineered features by approximately 4 AUROC points. On the SAP SALT enterprise benchmark, the model outperforms XGBoost with engineered features by 7 mean reciprocal rank (MRR) points out-of-the-box, rising to 13 MRR points after fine-tuning. Additionally, it surpasses AutoGluon ensembles by 8% MRR. The architecture supports horizontal scaling for databases containing over 500 billion rows.
Why it matters
Most existing foundation models for structured data operate on flattened tables, which discards fine-grained relational signals and demands manual aggregation. Kumo Relational processes data natively within relational structures, preserving foreign keys and timestamps. It employs a Predictive Query Language (PQL) that allows users and AI agents to define what to predict rather than how, streamlining the integration of predictive capabilities into agentic workflows.
The inference process involves three distinct mechanisms:
- Automatic Context Generation: For temporal tasks, the system replays database states by sampling entity-centric subgraphs up to an anchor timestamp. Targets are computed from post-anchor events while feeding the model only pre-anchor data, preventing information leakage.
- Within-Table Reasoning: A network alternates between column attention and row attention to correlate values and filter noise before propagation.
- Cross-Table Reasoning: Graph attention distributes row embeddings across primary-key and foreign-key relationships, interleaved with cross-sample attention. This approach avoids quadratic computational costs by operating on embeddings rather than individual cells.
Pre-training leverages synthetic data generated via PluRel, a tool developed with contributions from the Kumo team. PluRel creates schemas, clusters foreign keys, and assigns column values using Structural Causal Models with temporal trends. Pre-training loss follows a power law as synthetic volume increases, with performance gains transferring to real-world datasets.
What to do next
Developers can integrate Kumo Relational using NVIDIA’s open-source structured-data-models library or deploy it as an NVIDIA NIM. While the model functions effectively without training, organizations seeking maximum accuracy on specific workloads can perform fine-tuning. To initiate a query, users define the prediction goal in PQL, and the system handles context retrieval, subgraph extraction, and inference automatically.
Limitations
Kumo Relational was pre-trained entirely on synthetic data generated by PluRel. While the source indicates that pre-training losses follow a power law and gains transfer to real databases, reliance on simulated environments means validation against diverse, proprietary enterprise data distributions is essential before production deployment. The system is GPU-native, requiring compatible hardware for inference operations.
For related LockDraft context, see Ai2 Open-Sources AstaBrief 8B for Fast Scientific Report Generation.
Sources
Share Article
If this article helped you, please share it with others!