NVIDIA's Kumo Tabular is a one pass tabular model

NVIDIA released Kumo Tabular, an open foundation model for tabular classification and regression that skips the fitting stage entirely. You hand it a table and it returns predictions in a single forward pass, with no training, no tuning and no feature engineering. Three sizes are available, from 28M to 215M parameters, and the model was pretrained only on artificial data.

The claim that makes it interesting is the accuracy to efficiency ratio rather than raw score. On TabArena it ranks first overall at an ELO of 1950 while running 17 times faster than LimiX-2 under a uniform setup on a single RTX 6000 Pro. It also places first on BeyondArena at an ELO of 1418 with an improvability score of 7.78%, and tops the TALENT leaderboard overall, with average ranks of 6.67 for classification accuracy, 3.98 for log-loss and 4.22 for regression RMSE.

How it works

The architecture is a Transformer using column, row and in-context attention, building on TabICL and TabPFN. Numeric and categorical values get Fourier features, each row carries four learnable read out tokens, query rows are handled with grouped query attention, and attention temperature scales with the log of the number of keys so attention stays sharp on long tables. Classification and regression are trained as separate models across three stages, from 1,024 rows and up to 100 columns up to 60,000 rows, on roughly 35M, 71M and 137M artificial tables for the Small, Medium and Large variants. The regression head outputs 999 quantiles, which gives a point prediction plus an uncertainty estimate.

Weights are on Hugging Face and the code is on GitHub, under the OpenMDW License Agreement v1.1, which permits commercial use.

The limits are worth reading before planning around it. It handles numerical and categorical columns only, so text, images and timestamps need pre-processing recipes first. A single forward pass covers up to 10 classes, with error-correcting output codes needed beyond that. Accuracy degrades on tables outside the training ranges or when query rows come from a different distribution than the context rows, and NVIDIA’s own advice is to validate on your own held-out data. The training recipe and the artificial data generators are described as coming soon rather than shipping today.

(Original source: NVIDIA)