NVIDIA Kumo Relational Beats Supervised Learning Without Training
By Ali Al-Montaser — Senior AI Technical Editor
NVIDIA launches Kumo Relational, the first few-shot relational foundation model that outperforms supervised learning across 41 tasks in 15 databases. It beats AutoGluon by 8% in MRR on SAP SALT, runs without training or feature engineering, reads up to 10,000 in-context examples in a single forward pass, scales to databases exceeding 500 billion rows, and is available today under Apache 2.0 for code and OpenMDW 1.1 for weights.
NVIDIA Launches Kumo Relational: First Relational Foundation Model Outperforms Supervised Learning Without Training
NVIDIA has officially announced the availability of Kumo Relational today, a pre-trained relational foundation model that makes direct predictions on enterprise databases without task-specific training, feature engineering, or table flattening. The model is available immediately through the open-source structured-data-models library and as an NVIDIA NIM service, outperforming AutoGluon by 8% in MRR on the SAP SALT benchmark and running approximately 430 times faster than the top model on the RelArena-α leaderboard.
📊 Official Technical Specifications & Data Sheet
| Technical Dimension | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Free to use: code under Apache 2.0 and weights under OpenMDW 1.1; weights are downloaded from Hugging Face Hub on first use. No token pricing is announced — the model performs in-context inference on tabular data, not text token pricing. |
| 🌐 Platforms & Immediate Availability | Available today via: NVIDIA structured-data-models library (GPU-native, open source), NVIDIA NIM service, structured-data-models-client, GPU notebooks, and Databricks environment. Connectors for databases and repositories, plus a custom graph engine. |
| ⚡ Performance & Speed Benchmarks | +8% MRR vs. AutoGluon on SAP SALT (+16% after fine-tuning); +25% vs. modern tabular foundation models; +7 MRR points vs. out-of-the-box XGBoost and +13 after tuning; ~4 AUROC points on RelBench V1; 430x faster than the top model on RelArena-α (17 seconds vs. 7,300 seconds per 1,000 samples); Elo 1,842 on the accuracy-efficiency Pareto frontier. |
| 🛡️ Security & Breach Resistance | Prevents data leakage by replaying historical database states (pre-timestamp data only); stable when 75% of foreign key links are removed; stable performance with only 0.2% of training data as context; robust against noisy columns. |
| 🧠 Context Window | Up to 10,000 contextual examples in a single forward pass; scales to databases exceeding 500 billion rows via a memory-mapped I/O graph engine at speeds up to 5 GB/s. |
| 🌍 Arabic Language & Regional Support | The model specializes in relational and tabular data (numbers, categories, timestamps), not Arabic natural language processing; the PQL interface is in English, but it operates on databases with Arabic content without text token costs, making it practically available to organizations in the region. |
Announcement Details & New Technical Features
Kumo Relational operates directly on the database while respecting tables, foreign keys, and timestamps, instead of flattening relationships into a single row through manual aggregation that loses subtle structural signals. The model is pre-trained once on synthetic relational data that simulates real-world databases, then adapts to each new task via in-context learning by reading up to 10,000 examples, where each example is paired with a local subgraph around an entity at a point in time along with its true label (e.g., "this user churned" or "this item sold 142 units").
The architecture spans three stages: automatic context generation via the Predictive Query Language (PQL), which replays historical database states to prevent data leakage; then in-table inference with a lightweight network alternating between column attention and row attention to filter noise; then cross-table inference with a larger network that propagates row representations across the relational structure using graph attention over primary/foreign key relationships, interleaved with cross-sample attention. Because this stage operates on row representations rather than every cell, it avoids quadratic cost and scales to large contexts and deeper graphs.
The model is delivered through NVIDIA's new GPU-native structured-data-models library (the same library that runs its sibling Kumo Tabular), or as a standalone NVIDIA NIM service accessed by the structured-data-models-client from a notebook, making enterprise deployment straightforward.
Benchmark & Competitive Performance
Kumo Relational demonstrates consistent gains across multiple benchmarks. On SAP SALT, it achieves an 8% improvement in MRR over AutoGluon, rising to 16% after fine-tuning. Against modern tabular foundation models, it delivers a 25% improvement. Compared to out-of-the-box XGBoost, it gains 7 MRR points, and 13 points after tuning. On RelBench V1, it adds approximately 4 AUROC points. Speed is a major differentiator: it is 430 times faster than the top model on RelArena-α, completing 1,000 samples in 17 seconds versus 7,300 seconds. Its Elo score of 1,842 on the accuracy-efficiency Pareto frontier underscores its balanced performance.
Industry Impact & Enterprise Adoption
Kumo Relational addresses a critical pain point in enterprise AI: the need for specialized training and feature engineering for every predictive task. By eliminating these requirements, it enables organizations to deploy predictive analytics directly on relational databases with minimal overhead. The model's ability to scale to over 500 billion rows and maintain stability even with 75% of foreign key links removed makes it suitable for large-scale, real-world enterprise data. Its availability under permissive open-source licenses (Apache 2.0 for code, OpenMDW 1.1 for weights) lowers barriers to adoption, and its integration with NVIDIA NIM and Databricks facilitates deployment in existing workflows. For Arabic-speaking enterprises, the model's language-agnostic handling of numeric and categorical data means it can be applied to Arabic-content databases without additional text processing costs.
Conclusion
NVIDIA Kumo Relational represents a significant leap in relational foundation models, combining state-of-the-art performance with practical usability. Its few-shot learning capabilities, impressive speed, and open licensing position it as a transformative tool for enterprises looking to leverage AI on relational data without the traditional overhead of training and feature engineering. As organizations increasingly seek to extract value from their databases, Kumo Relational offers a compelling, efficient, and accessible solution.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
NVIDIA Kumo Relational is a pre-trained relational foundation model that makes direct predictions on enterprise databases without task-specific training. It outperforms AutoGluon by 8% in MRR on SAP SALT, beats XGBoost by 7 MRR points out-of-the-box and 13 points after fine-tuning, and is approximately 430 times faster than the top model on the RelArena-α leaderboard (17 seconds versus 7,300 seconds per 1,000 samples).
No. It relies on in-context learning by reading up to 10,000 contextual examples in a single forward pass, without task-specific training, feature engineering, or table flattening. Fine-tuning is optional only when maximum accuracy is needed.
It is available today through NVIDIA's open-source GPU-native structured-data-models library (which downloads weights from Hugging Face Hub on first use) or as an NVIDIA NIM service with the structured-data-models-client. The code is under Apache 2.0 and the weights under OpenMDW 1.1.
It scales to databases exceeding 500 billion rows through a custom graph engine based on memory-mapped I/O at speeds up to 5 GB/s, while maintaining stable performance even when 75% of foreign key links are removed.
The model specializes in tabular and relational data (numbers, categories, timestamps) rather than natural language processing of Arabic. The PQL query interface is in English, but it operates on databases with Arabic content because numeric and categorical values are language-agnostic, making it practically usable for organizations in the region without text token costs.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
