NVIDIA Kumo brings foundation models to ordinary business tables
The September 29 release predicts classifications and numeric values from labeled rows, with three model sizes and open weights under OpenMDW-1.1.
What happened
NVIDIA introduced Kumo Tabular on September 29 as a foundation-model family for classification and regression on tables. Given labeled rows as context, the model predicts new rows in a forward pass without task-specific training. Three sizes span 28 million to 215 million parameters.
Where the results come from
NVIDIA reports leading results on TabArena, BeyondArena, TALENT and ScoringBench. It says pretraining used artificial tables and provides weights plus an open-source inference library. AITrending has checked the release and model metadata, but has not reproduced those benchmark results.
What to test on business data
The core model handles numerical and categorical columns; text, images and timestamps require preprocessing into features. NVIDIA warns that performance can fall when tables exceed the training ranges or when query rows differ in distribution from the context.
The model card lists OpenMDW-1.1. For a real deployment, the next step is held-out evaluation and calibration on the intended data, including comparison with existing tree-based models. A public leaderboard result does not establish the best choice for a particular company’s table.