NVIDIA has entered the tabular foundation model space with Kumo Tabular, an open-source model that eliminates traditional barriers to enterprise machine learning on structured data. Released today on Hugging Face, the model delivers predictions on tabular data without the usual requirements for training, feature engineering, or hyperparameter tuning—a significant shift from the gradient-boosted tree models that have dominated enterprise analytics for two decades. Users simply provide a labeled table and receive predictions on new rows in a single forward pass, supporting both classification and regression across tables of any size. The model architecture addresses tabular data through specialized Transformer mechanisms designed explicitly for tables. Three types of attention—column, row, and in-context—work together to extract meaning from individual cell values, learn feature interactions, and apply context-based predictions. A novel length-aware attention temperature prevents performance degradation as table size scales up or down, ensuring consistent accuracy whether working with hundreds or tens of thousands of rows. Trained entirely on synthetic data generated through Structural Causal Models, the model comes in three sizes from 28M to 215M parameters. Kumo Tabular's performance across industry benchmarks underscores its competitive positioning. The model ranks first place on all four major evaluation suites—TabArena, BeyondArena, TALENT, and ScoringBench. The commercial release under the OpenMDW-1.1 license, combined with open-source code availability, signals NVIDIA's commitment to making tabular foundation models accessible to enterprises seeking faster analytics pipelines.