Particle.news
Download on the App Store

KumoRFM-2 Brings Plain-English Predictions to Enterprise Relational Data

It promises to skip ETL and task-specific training, with results that still need outside verification.

Overview

  • Kumo introduced a foundation model built to reason over linked tables in enterprise databases used for day-to-day business data.
  • The model keeps foreign-key relationships intact and processes multiple connected tables without manual flattening into a single sheet.
  • Users can run plain-English predictive queries against Snowflake, Databricks, or other SQL warehouses with no feature store, ETL, or task-specific training.
  • Company materials and an arXiv paper report stronger results than supervised baselines on Stanford RelBench and SAP SALT, plus resilience to noisy or missing data and cold starts.
  • Kumo says the system scales from billion-row datasets to hundreds of billions of rows, a claim that enters a crowded field and still needs independent checks against rivals like SAP-RPT-1, MotherNet, TabICL, and AWS Mitra.