NVIDIA Kumo Tabular: Open Tabular Foundation Models for Prediction
Researchers have introduced Kumo Tabular, a family of open tabular foundation models designed for classification and regression across three distinct sizes ranging from 27M to 214M parameters. Pretrained exclusively on artificial tables from the Kumo prior, these models can make predictions on new tables without additional training. Kumo Tabular-L ranks first across eight benchmarks and demonstrates significantly faster prediction speeds than comparable systems. Open article →