NVIDIA releases Kumo Tabular, an open foundation model for tabular prediction on Hugging Face

NVIDIA has released Kumo Tabular, an open foundation model for tabular prediction, hosted on Hugging Face. The model supports both classification and regression and operates without training, tuning, or feature engineering, predicting labels for new rows in a single forward pass given a labeled table as context.

Kumo Tabular is part of NVIDIA’s Kumo Structured model collection and is released under the OpenMDW-1.1 license for commercial use. It is available in three sizes ranging from 28M to 215M parameters and is accompanied by an open-source library and model weights on GitHub and Hugging Face.

The model architecture is a Transformer that uses column, row, and in-context attention to interpret tables. It embeds cells with Fourier features, handles missing values explicitly, and uses row and column attention to capture feature interactions and column-wise distributions. In-context learning enables the model to condition predictions on labeled context rows without updating weights.

Kumo Tabular was pretrained entirely on artificial tables generated from structural causal models, including mechanisms for missingness, outliers, and heavy-tailed targets. Training proceeds in three stages with increasing table sizes and context lengths, and the model is trained separately for classification and regression.

According to the announcement, Kumo Tabular ranks first on four benchmarks: TabArena, BeyondArena, TALENT, and ScoringBench. The model is positioned as a shift from traditional gradient-boosted trees toward a foundation-model approach for tabular prediction.