NVIDIA has open-sourced Kumo Relational, a foundation model that predicts outcomes directly from enterprise relational databases — and beats XGBoost, the decade-long gold standard for tabular machine learning, across 41 benchmark tasks without any manual feature engineering.
Published on Hugging Face, the model uses in-context learning: it accepts up to 10,000 historical database records as context in a single forward pass and returns predictions without retraining. What previously took weeks of feature engineering — flattening multi-table databases and hand-crafting columns — now reduces to a declarative query.
At its core is the Relational Graph Transformer, a three-stage architecture that reasons natively over primary-key and foreign-key relationships instead of first converting tables into flat representations. NVIDIA reports the model scales to 500 billion database rows. The release grew out of NVIDIA’s reported $400 million acquisition of Kumo AI in June 2026, the startup co-founded by Stanford professor Jure Leskovec.
On benchmarks spanning e-commerce, medical, enterprise, social, and academic databases, Kumo Relational beats gradient-boosted tree pipelines trained on expert-engineered features by nearly four AUROC points on RelBench V1 classification tasks, and outperforms large AutoGluon ensembles on SAP’s real enterprise ERP data by 8% without fine-tuning — extending to 16% with fine-tuning.
The model is available under an Apache 2.0 library and weights on Hugging Face under NVIDIA’s OpenMDW license, which permits commercial use with restrictions. For agentic AI stacks built on models like Claude and ChatGPT, it offers a path to anticipating outcomes — churn, delays, fraud — directly from enterprise data.
Source: Tech Times
