NVIDIA has released Kumo Tabular, an open foundation model for tabular classification and regression, as part of the Kumo Structured model collection. It is available on Hugging Face under the OpenMDW-1.1 license and runs through the open-source structured-data-models library. Given a table of labeled rows, the model predicts labels for new rows in a single forward pass, with no training, no tuning, and no feature engineering. Three sizes are available, ranging from 28M to 215M parameters, and NVIDIA reports it ranks first on TabArena, BeyondArena, TALENT, and ScoringBench.

Kumo Tabular is a Transformer built around table structure, combining column, row, and in-context attention as introduced in TabICL and TabPFN. Cells are grouped into tokens; numerical and categorical values pass through Fourier features with learned frequencies, with separate weights for each type. Missing values are treated specially and need no imputation, and every context token receives a label embedding. Rows are compressed by alternating two attention kinds: column attention looks down a single column to learn what a value means in its distribution, with cost growing linearly in the number of rows; row attention looks across the tokens of a single row to learn feature interactions, using rotary positions to distinguish columns. Four learnable [CLS] tokens join each row and act as the final row readout, so the final stage no longer depends on column count. A final Transformer then operates on the row embeddings: context rows attend to each other, while query rows attend only to context rows, making each prediction independent of which other rows are scored alongside it. Query rows use Test-GQA to shrink the prediction cache, and a head produces class probabilities for classification or 999 quantiles for regression, from which a point prediction and uncertainty estimate follow.

Because softmax attention spreads out as the number of keys grows, Kumo Tabular scales every query by a temperature that grows with the logarithm of the number of keys, with a coefficient learned separately for each attention head. This length-aware attention temperature keeps attention sharp as tables grow longer or wider than typical training tables.

The model is pretrained entirely on artificial tables. Each training table is sampled from a Structural Causal Model in several steps: a configuration is drawn for table size, task, mechanisms, and missingness; a random causal graph links hidden variables evaluated from root to leaf through randomly drawn functions such as linear maps, small neural networks, trees, or Gaussian processes; some nodes become numerical or categorical columns, one becomes the target, and the rest stay hidden like unmeasured causes. Post-processing correlates groups of columns, clips outliers, and injects missing values, and a quick tree-ensemble check discards tables without a learnable signal. Because the generator is a procedural sampler rather than a trained model, it produces an endless supply of new tables. Real-world messiness is built in: multiple missingness patterns, coarsened features that can make duplicate rows disagree on labels, categorical columns with many levels, and heavy-tailed regression targets. On every table, the model sees most rows with labels as context and learns to predict the remaining rows, using cross-entropy loss for classification and quantile loss for regression; classification and regression are trained as separate models. Training runs in three stages: the first uses tables of 1,024 rows and up to 100 columns, the second varies context from 400 to 10,240 rows, and the third extends context to 60,000 rows. In total, Kumo Tabular-Small/Medium/Large saw about 35, 71, and 137 million artificial tables. The training recipe and artificial data generators are planned for release soon.

On the full TabArena leaderboard, with default settings against tuned gradient-boosted trees, AutoGluon, and other tabular foundation models, Kumo Tabular ranks first overall with an ELO of 1950 while running 17x faster than LimiX-2 under a uniform single RTX 6000 Pro evaluation setup. Across all three model sizes, NVIDIA reports a new state of the art on the accuracy-efficiency Pareto front. On BeyondArena, Kumo Tabular reaches an ELO of 1418 with an Improvability score of 7.78%, placing first. On TALENT, it achieves 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.

Kumo Tabular works on numerical and categorical columns only; text, images, or timestamps can be turned into features via built-in preprocessing recipes. A single forward pass covers up to 10 classes, which the library extends to any number of classes with error-correcting output codes. Accuracy may degrade on tables far beyond the training ranges or when query rows come from a different distribution than the context rows, so NVIDIA recommends validating accuracy and calibration on held-out data before deployment. The structured-data-models library downloads weights from the Hub on first use and provides the preprocessing, ensembling, and many-class handling used in the evaluations; the source shows a path from a pandas DataFrame to predictions via sdm.TableTensor.from_pandas and sdm.models.KumoTabular.