NVIDIA introduced Kumo Tabular, a downloadable foundation model for tabular prediction that supports classification and regression without task-specific training.
NVIDIA introduced Kumo Tabular on September 29, 2026, as an open foundation model for classification and regression on tabular data. Published on Hugging Face, the model takes labeled rows as context and predicts labels for new rows in a forward pass, without task-specific training, tuning, or feature engineering, according to NVIDIA.
The release includes model sizes ranging from 28 million to 215 million parameters and uses an open-source library for execution. NVIDIA says the models were pretrained on artificial tables generated from structural causal models rather than on real-world tables. The company reports that the small, medium, and large versions were trained on about 35 million, 71 million, and 137 million artificial tables, respectively. NVIDIA says it plans to release the training recipe and artificial-data generators separately.
The model’s architecture uses a Transformer designed around rows and columns. NVIDIA describes distinct attention mechanisms for interpreting values within columns, learning how features interact within rows, and using labeled examples to make predictions for query rows. It supports numerical and categorical columns; text, images, and timestamps can be represented through preprocessing recipes. The model handles up to 10 classes in a single forward pass, while the library supports larger class counts through error-correcting output codes.
NVIDIA reports top placements across four evaluations: TabArena, BeyondArena, TALENT, and ScoringBench. On TabArena, the company reports an overall ELO of 1950. On BeyondArena, it reports an ELO of 1418 and an Improvability score of 7.78%. These are results reported by NVIDIA; the post also advises users to validate accuracy and calibration on held-out data before deployment.
The model is released under the OpenMDW License Agreement, version 1.1, which NVIDIA says permits commercial use under its terms. The post cautions that performance may weaken on tables well outside the training ranges or when prediction rows differ in distribution from the labeled context. Those limits make testing on the intended dataset an important step before using the model in production.