Tabular data remains the workhorse of enterprise machine learning, but most models still learn each task from scratch. NVIDIA's Kumo Tabular, now available on Hugging Face, takes a different approach: given a table with labeled rows, it predicts labels for new rows in a single forward pass, with no training, tuning, or feature engineering. The model handles both classification and regression and comes in three sizes, from 28M to 215M parameters.
Kumo Tabular is a Transformer that combines column, row, and in-context attention. It embeds cells using Fourier features, treats missing values without imputation, and compresses rows before a final attention stage relates labeled context rows to query rows. A length-aware attention temperature keeps predictions sharp even when inference tables are much larger than training tables.
The model was pretrained entirely on artificial tables generated from structural causal models, and it ranks first on the TabArena, BeyondArena, TALENT, and ScoringBench benchmarks. It is released under the OpenMDW-1.1 license for commercial use, with code and weights available on GitHub and Hugging Face.