Nvidia has released Kumo Tabular, an open foundation model for tabular data that predicts the labels of new rows in a single forward pass without training, tuning or feature engineering. The model, part of the Nvidia Kumo Structured collection, is available on Hugging Face under the OpenMDW-1.1 license for commercial use. It matters because tabular prediction — churn, default, demand and price — has for two decades been dominated by gradient-boosted trees that must be rebuilt from scratch for every new question, and Nvidia claims Kumo Tabular replaces that lifecycle with in-context learning.
Kumo Tabular comes in three sizes, from 28 million to 215 million parameters, and handles both classification and regression. It was pretrained only on artificial data. Given a table of labeled rows, it returns class probabilities or numeric predictions directly; for regression it produces 999 quantiles, from which a point prediction and an uncertainty estimate follow. The model runs through Nvidia’s open-source library, with weights on Hugging Face and code on GitHub.
The architecture is a Transformer built around table structure, using column, row and in-context attention as introduced in TabICL and TabPFN. Cells become tokens through Fourier features, with separate weights for numerical and categorical values and special handling for missing values that requires no imputation. Column attention learns what a value means within its column at a cost that grows linearly with rows; row attention learns how features interact, with four learnable [CLS] tokens acting as the final readout. A final Transformer operates on row embeddings, where context rows attend to each other and query rows attend only to context, so each prediction depends only on the context and the row itself. A length-aware attention temperature scales every query by a temperature that grows with the logarithm of the number of keys, keeping attention sharp as tables grow longer or wider.
Training used entirely artificial tables sampled from a Structural Causal Model. Each table is drawn from a random causal graph evaluated from root to leaf, with some nodes becoming numerical or categorical columns, one becoming the target, and the rest staying hidden. Post-processing correlates column groups, clips outliers and injects missing values, and a tree-ensemble check discards tables without a learnable signal. Training ran in three stages: the first used tables of 1,024 rows and up to 100 columns, the second varied context from 400 to 10,240 rows, and the third extended it to 60,000 rows. Kumo Tabular-Small, Medium and Large saw about 35, 71 and 137 million artificial tables respectively. Classification and regression are trained as separate models, and Nvidia says its training recipe and artificial data generators will be released soon.
Nvidia reports that Kumo Tabular ranks first overall on the full TabArena leaderboard with an ELO of 1950, running 17 times faster than LimiX-2 under a uniform single RTX 6000 Pro evaluation setup. On BeyondArena it reaches an ELO of 1418 with an Improvability score of 7.78%, placing first. On TALENT it takes the top overall ranking across classification accuracy, classification log-loss and regression RMSE, with average ranks of 6.67, 3.98 and 4.22. On ScoringBench, a benchmark for predictive distributions, Kumo Tabular-Large and Medium rank first and second on average rank.
Several limitations remain. Kumo Tabular works on numerical and categorical columns only; text, images or timestamps must be turned into features through built-in pre-processing recipes. A single forward pass covers up to 10 classes, which the library extends to any number of classes with error-correcting output codes. Nvidia also warns that accuracy may degrade on tables far beyond the training range. The benchmarks are Nvidia’s own runs, and the training recipe and data generators have not yet been released.
