Kumo Unveils KumoRFM-2, a Foundation Model Scaling to 500 Billion Rows

Kumo has launched KumoRFM-2, a foundation model that surpasses fully supervised machine learning on enterprise relational data, announced in a press release. The model is designed to eliminate the need for manual feature engineering and custom model training, enabling users to query data in plain English.

KumoRFM-2 is built on a new Relational Graph Transformer architecture and can process data at 5 GB per second, handling over 20 million lookups each second. It operates directly on connected tables without flattening and scales to more than 500 billion rows.

On benchmarks such as Stanford RelBenchV1, KumoRFM-2 outperformed its predecessor by 10% and exceeded top supervised machine learning models by 5% in classification and regression tasks. It also achieved state-of-the-art results on the SAP SALT benchmark, surpassing tabular model ensembles and other foundation models. Fine-tuning further improved performance by 13%, with the model demonstrating robustness to noise, missing data, and structural inconsistencies.