NVIDIA Kumo Tabular Predicts a Table's Labels in One Forward Pass
NVIDIA's 29 September 2026 post says the open tabular model needs no training step and ships in sizes from 28 million to 215 million parameters.
Direct answer
NVIDIA says Kumo Tabular, published 29 September 2026, predicts labels for new table rows in one forward pass, was pretrained only on artificial tables, and is released under OpenMDW-1.1.
NVIDIA's Hugging Face post on 29 September 2026 introduces Kumo Tabular as an open foundation model for tables. Given labeled rows and the rows to score, the post says it returns class probabilities or numeric predictions in a single forward pass, with no training, no tuning, and no feature engineering. Classification and regression are both in scope. Pretraining used only artificial tables. The weights come in three sizes, from 28 million to 215 million parameters, and the license named in the post is OpenMDW-1.1.
The post says the model ranks first on TabArena, BeyondArena, TALENT, and ScoringBench. On TabArena it reports an ELO of 1950. On BeyondArena it reports an ELO of 1418 and an improvability score of 7.78 percent. The code link is github.com/NVIDIA/structured-data-models and the weights link is huggingface.co/nvidia/Kumo-Tabular.
What the post says the model will not do raw
The limitations section limits raw input to numerical and categorical columns. Text, images, and timestamps go through preprocessing recipes. One forward pass covers up to 10 classes, and the library extends that with error-correcting codes. The post also says accuracy can drop when a table is far outside the training range, or when query rows come from a different distribution than the context rows, and it tells readers to check a held-out set before deployment.
Source: NVIDIA, Kumo Tabular.
FAQ
- Which benchmarks does the post say it leads?
- The post says it ranks first on TabArena, BeyondArena, TALENT, and ScoringBench, and gives a TabArena ELO of 1950.
- What does it not take as raw columns?
- The limitations section says it works on numerical and categorical columns. Text, images, or timestamps need the library's preprocessing.