NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction

NVIDIA Kumo Tabular is a new open foundation model designed for tabular data prediction. It performs classification and regression on new data rows in a single forward pass using in-context learning, eliminating the need for training, tuning, or feature engineering. Pretrained exclusively on artificial data, the model is available in three sizes and is built upon a Transformer architecture that utilizes column, row, and in-context attention. Kumo Tabular aims to provide a more efficient alternative to traditional gradient-boosted trees for enterprise machine learning tasks. It is released for commercial use and is available through an open-source library on Hugging Face.