NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction
NVIDIA introduces Kumo Tabular, an open foundation model for tabular data prediction that requires no training or feature engineering, outperforming existing methods across four benchmarks.
NVIDIA Kumo Tabular, an open foundation model for tabular data, is now available on Hugging Face as part of the NVIDIA Kumo Structured model collection. The model predicts labels for new rows in a single forward pass without training, tuning, or feature engineering, handling both classification and regression tasks. It is pretrained exclusively on artificial data, released under the OpenMDW-1.1 license for commercial use, and comes in three sizes ranging from 28M to 215M parameters.
Tabular data underpins many enterprise machine learning applications, including customer records, transactions, and sensor logs, where predicting outcomes like churn or demand is common. Traditional approaches rely on gradient-boosted trees, requiring extensive feature engineering and model retraining for each new task. Kumo Tabular adopts an in-context learning approach, enabling it to generalize from labeled tables without updating weights, similar to how large language models handle new tasks with minimal input.
Kumo Tabular is a Transformer model designed specifically for tabular data, utilizing column, row, and in-context attention mechanisms to understand column interactions, row relationships, and label dependencies. It processes numerical and categorical values through Fourier features, handles missing values without imputation, and employs length-aware attention to maintain performance as table sizes grow. The model is pretrained on millions of procedurally generated artificial tables, each sampled from a Structural Causal Model to simulate real-world imperfections.
In evaluations across four benchmarks—TabArena, BeyondArena, TALENT, and ScoringBench—Kumo Tabular ranks first overall, achieving an ELO of 1950 on TabArena and outperforming tuned gradient-boosted trees, AutoGluon, and other tabular foundation models. On BeyondArena, it reaches an ELO of 1418 with a 7.78% improvability score, while on TALENT, it leads in classification accuracy, log-loss, and regression RMSE. The model is available in three sizes and runs efficiently on a single RTX 6000 Pro GPU.