Open model targets faster, training-free prediction on tabular data
NVIDIA has released Kumo Tabular, an open foundation model for classification and regression on tabular data. Available through Hugging Face, it accepts rows with known labels as context and predicts labels for new rows in a single forward pass, without task-specific training, tuning or feature engineering. The model is offered in three sizes ranging from 28M to 215M parameters and is licensed under OpenMDW-1.1 for commercial use.
Kumo Tabular is a Transformer designed around column, row and in-context attention. It was pretrained entirely on procedurally generated tables that include missing values, categorical features, outliers and other imperfections. NVIDIA trained separate classification and regression models and says the largest version was exposed to about 137 million artificial tables.
In NVIDIA’s evaluations, Kumo Tabular ranked first on TabArena, BeyondArena, TALENT and ScoringBench, including the top overall TabArena result with an ELO of 1950. The model handles numerical and categorical columns directly; text, images and timestamps require preprocessing. NVIDIA notes that performance can decline when data falls far outside training ranges or query rows differ from context rows, and recommends checking accuracy and calibration on held-out data before deployment.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.