NVIDIA’s Kumo Tabular Predicts From Labeled Rows Without Retraining or Feature Engineering

NVIDIA published Kumo Tabular on September 29, 2026, as an open foundation model for tabular classification and regression that returns class probabilities or numeric predictions in a single forward pass—without task-specific training, hyperparameter search, or hand-built features. The drop sits in the NVIDIA Kumo Structured model collection on Hugging Face, ships with a GPU-native open-source library on GitHub (NVIDIA/structured-data-models), and is licensed under the OpenMDW License Agreement version 1.1 for commercial use. Three sizes span roughly 28 million to 215 million parameters. NVIDIA’s evaluation claim is blunt: with default settings, Kumo Tabular ranks first on TabArena, BeyondArena, TALENT, and ScoringBench under the company’s reported setups.

Enterprise machine learning still lives mostly in tables—churn, default, demand, claims, sensor logs—and for two decades that work has leaned on gradient-boosted trees. Those models work, but each new question restarts a familiar cycle: collect labels, engineer features, tune, validate, deploy something that knows nothing about tables in general. Large language models popularized in-context learning for text; Kumo Tabular applies the same idea to tables. Given labeled context rows and unlabeled query rows, the model reads the labeled table as context and predicts the missing labels directly.

Architecturally it is a Transformer built around table structure, using column, row, and in-context attention in the lineage of TabICL and TabPFN. Cell embedding turns groups of cells into tokens: numerical and categorical values go through Fourier features with separate weights, missing values are handled without imputation, and context tokens receive label embeddings. Row embedding alternates column attention—linear in the number of rows via induced self-attention that learns what a value means in its column—and row attention across a row’s tokens with rotary positions, plus four learnable [CLS] tokens as the row readout. A final Transformer then runs in-context learning: context rows attend to each other, while query rows attend only to context, so predictions do not depend on which other queries share the batch. Context keys and values can be computed once and reused. Query rows use Test-GQA to shrink the cache each prediction reads. Classification heads emit class probabilities; regression heads emit 999 quantiles from which a point estimate and uncertainty follow. Softmax attention also gets a length-aware temperature that grows with the log of the number of keys, with a learned coefficient per head, so attention stays sharp as tables grow longer or wider than typical training tables.

Training is entirely on artificial tables sampled from Structural Causal Models. A procedural generator draws table size, task, mechanisms, and missingness; builds a random causal graph; evaluates random functions at nodes; exposes some nodes as columns and one as the target; then injects missingness patterns, coarsening, high-cardinality categoricals, and heavy-tailed regression targets. Tables without a learnable signal are discarded by a quick tree-ensemble check. Classification and regression train as separate models across three stages that expand context from 1,024 rows up to 60,000 rows and up to 100 columns. NVIDIA reports that Small, Medium, and Large saw about 35 million, 71 million, and 137 million artificial tables respectively. The training recipe and generators are promised for a later release.

On performance, NVIDIA says Kumo Tabular posts an overall TabArena ELO of 1950 while running 17× faster than LimiX-2 under a uniform single RTX 6000 Pro setup, with all three sizes on a new accuracy–efficiency Pareto front against tuned gradient-boosted trees, AutoGluon, and other tabular foundation models. BeyondArena lists an ELO of 1418 (7.78% Improvability) for first place; TALENT leads overall on classification accuracy, log-loss, and regression RMSE; ScoringBench puts Large and Medium first and second on average rank for predictive distributions. Limits are stated clearly: only numerical and categorical columns in the core path (text, images, or timestamps need preprocessing recipes); a single forward pass covers up to 10 classes, with the library extending further via error-correcting output codes; accuracy can degrade far outside training ranges or under distribution shift between context and query rows.

Day-to-day use goes through the structured-data-models library (import sdm), which downloads Hub weights on first use and wraps preprocessing, ensembling, and many-class handling. A short path tensorizes a pandas DataFrame to a TableTensor on CUDA, masks unlabeled targets, and calls sdm.models.KumoTabular with context features/labels and query features. The GitHub library itself is marked Apache 2.0 for contributions and packaging; model weights remain under OpenMDW-1.1.

For Writeble’s opensource desk, Kumo Tabular matters because it pushes foundation-model convenience into the tabular workloads that still dominate enterprise ML—and it ships weights, code, and hard limits teams can verify.