NVIDIA's KumoRFM, a relational foundation model, outperforms supervised machine learning on relational data without requiring any task-specific training. In benchmarks across approximately 100 tables and 2 billion rows spanning 15 real-world relational graphs, KumoRFM scores 76.71 average AUROC zero-shot across 12 entity-classification tasks, ahead of LightGBM at 62.44, an LLM baseline at 68.06, and supervised relational deep learning at 75.83. Fine-tuning raises the average to 81.14 AUROC. The New Stack, referring to the model as KumoRFM-2, reported it accepts plain-English queries and replaces what previously required months of data science engineering work.
The speed difference is substantial: zero-shot inference takes approximately one second and one line of PQL versus 30 minutes and 56 lines for a relational GNN pipeline and 12.3 hours and 878 lines for a manual data-scientist workflow. The model is pretrained on diverse relational datasets so its parameters learn general relational patterns; during zero-shot inference those parameters stay frozen while the new database, target definition, and historical labeled context are supplied dynamically. It works with declared schemas, entity tables, fact tables, and relationship links without requiring callers to flatten connected tables into a single feature table.
The Kumo family covers relational and tabular prediction, forecasting, and anomaly detection. Developers can use pretrained Kumo models and fine-tune them with proprietary data for higher performance. Early adopters include Cognite, Coinbase, Databricks, DoorDash, Revolut, and Tubi, and Snowflake customers in advertising, energy, financial services, and retail are already using Kumo. At GTC Berlin on October 21, 2026, Jensen Huang positioned Kumo within the Agent Knowledge Foundation, describing it as providing agents with access to company knowledge. Dozens of partners including Palantir, Databricks, Snowflake, AWS, IBM, Oracle, and Zoom are building on or integrating with the Agent Knowledge Foundation.