Kumo’s new foundation model replaces months of data science engineering with plain-English queries

Kumo has announced KumoRFM-2, a foundation model purpose-built for enterprise relational data that can be queried in plain English with zero training. Unlike LLMs trained on unstructured text, it natively reasons over multi-table relational structures by working directly on the graph of connected tables, preserving foreign-key relationships. On the Stanford RelBench v1 benchmark, it outperforms its predecessor by 10% and beats the strongest supervised ML model by 5%. It connects directly to data warehouses like Snowflake or Databricks with no ETL, feature stores, or model training required, and scales to over 500 billion rows. The post also surveys competing tabular foundation models including SAP-RPT-1, MotherNet, TabICL, and Amazon's Mitra.