NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction
huggingface ·
NVIDIA has released Kumo Tabular, an open foundation model designed for tabular data prediction that operates without training or feature engineering. This Transformer-based architecture utilizes in-context learning to predict labels for new rows in a single forward pass, supporting both classification and regression tasks. The model was pretrained exclusively on artificial tables generated by Structural Causal Models, allowing it to handle various data imperfections like missing values and outliers. It is available in three sizes ranging from 28 million to 215 million parameters and runs via NVIDIA's open-source structured-data-models library. Performance evaluations show that Kumo Tabular ranks first on major benchmarks including TabArena, BeyondArena, TALENT, and ScoringBench. Notably, it establishes a new state-of-the-art on the accuracy-efficiency Pareto front while running significantly faster than competing models like LimiX-2. Although currently limited to numerical and categorical columns, the model offers robust handling of complex table structures through specialized attention mechanisms. Released under the OpenMDW-1.1 license, it provides developers with immediate access to high-performance tabular AI capabilities.
huggingface