NVIDIA Releases Kumo Tabular for Training-Free Table Predictions

Published on September 29, 2026, NVIDIA Kumo Tabular is an open foundation model for tabular data available on Hugging Face as part of the NVIDIA Kumo Structured model collection. [1] Given labeled rows and rows requiring predictions, it produces class probabilities or numeric predictions for classification and regression in one forward pass, without training, tuning, or feature engineering. [1]

It was pretrained exclusively on artificial data, comes in three sizes spanning 28 million to 215 million parameters, and runs through an open-source library. [1] Released under OpenMDW-1.1 for commercial use, the model ranks first on TabArena, BeyondArena, TALENT, and ScoringBench, according to NVIDIA’s Hugging Face post. [1]

Kumo Tabular packages tabular classification and regression into an in-context workflow: labeled rows provide context, while the pretrained model returns predictions without updating weights. That contrasts with the source’s description of the conventional gradient-boosted-tree lifecycle, where each question requires labels, feature engineering, hyperparameter search, validation, and deployment. Its three model sizes and open-source library give evaluators multiple deployment options, while first-place results across four named benchmarks establish the performance claim presented by NVIDIA. What the sources don't address: how Kumo Tabular performs on organizations’ private, noisy, or distribution-shifted tables outside the four cited benchmarks.