NVIDIA Kumo Tabular Launch
NVIDIA Kumo brings foundation models to ordinary business tables
28M to 215M parameters Verified 2 sources
N
NVIDIA Kumo Tabular
Model Sep 30, 2026
NVIDIA's tabular foundation models predict classifications or numeric values from labeled examples, with three sizes spanning 28M to 215M parameters.
Visit primary link
2 sources attached
What happened lately
NVIDIA Kumo Tabular timeline
What it does
NVIDIA announced Kumo Tabular on September 29. Given labeled rows as context, it predicts labels or numeric values for new rows in a forward pass, covering classification and regression. This is a model for structured tables rather than a conversational assistant.
The release has three sizes, from 28 million to 215 million parameters. NVIDIA says it pretrained the models entirely on artificial tables and provides an open-source structured-data-models library for inference and preprocessing.
Results and limits
NVIDIA reports leading results on TabArena, BeyondArena, TALENT and ScoringBench. Those are the release team’s benchmark claims, not independently reproduced AITrending results.
The core model handles numerical and categorical columns. Text, images and timestamps require feature preprocessing. The release notes warn that accuracy can degrade when data lies far outside the training ranges or when prediction rows differ in distribution from the labeled context.
Access
The model card declares the OpenMDW-1.1 license. Validate accuracy and calibration on held-out business data before deployment. The Hub reported 49 likes and zero recorded downloads on September 30; a zero counter does not establish that nobody has used the release.