NVIDIA Kumo Tabular Tops Four Benchmarks With Zero Training Predictions
NVIDIA releases Kumo Tabular, an open model that tops four benchmarks, predicts table data in one pass, and allows commercial use through Hugging Face.
Summary
On September 29, 2026, NVIDIA released Kumo Tabular, an open foundation model in its Kumo Structured collection. From labeled rows, it produces class probabilities or numeric predictions in one forward pass, without task training, tuning, feature engineering, or missing value imputation. Separate classification and regression models come in 28 million to 215 million parameter sizes, run through NVIDIA's GPU native structured data models library, download weights from Hugging Face, and permit commercial use under OpenMDW 1.1.
The Transformer combines cell, column, row, and in-context attention derived from TabICL and TabPFN, caches context computations, uses Test-GQA to reduce cache size, and adjusts attention temperature as tables grow. Regression returns 999 quantiles for point and uncertainty estimates. Pretraining used only procedurally generated structural causal model tables, including missingness, high cardinality categories, duplicate conflicts, and heavy tailed targets. Small, Medium, and Large saw about 35 million, 71 million, and 137 million tables. Three stages progressed from 1,024 rows to contexts of 400 to 10,240, then 60,000 rows, always up to 100 columns. NVIDIA plans to release the generators and recipe soon.
Under default settings, Kumo Tabular led TabArena at 1950 ELO and ran 17 times faster than LimiX-2 on a single RTX 6000 Pro, against tuned gradient boosted trees, AutoGluon, and foundation models. It led BeyondArena at 1418 ELO and 7.78% Improvability; TALENT with average ranks of 6.67 for classification accuracy, 3.98 for classification log loss, and 4.22 for regression RMSE; and ScoringBench, where Large and Medium ranked first and second. Native columns are numerical or categorical; recipes convert text, images, and timestamps. One pass handles 10 classes, with error correcting output codes extending that limit. Accuracy can fall beyond training ranges or when query and context distributions differ, requiring held out validation of accuracy and calibration.
Positives
- Kumo Tabular ranks first across TabArena, BeyondArena, TALENT, and ScoringBench evaluations.
- TabArena performance reached 1950 ELO while running 17 times faster than LimiX-2 on one RTX 6000 Pro.
- Three model sizes from 28 million to 215 million parameters make different accuracy and efficiency tradeoffs available.
- OpenMDW 1.1 permits commercial use, while code and weights are available through GitHub and Hugging Face.
- One forward pass produces classification or regression predictions without task training, tuning, feature engineering, or missing value imputation.
Risks & concerns
- Native inputs support only numerical and categorical columns, requiring preprocessing recipes for text, images, and timestamps.
- One forward pass directly supports only 10 classes, although error correcting output codes extend the library to larger class counts.
- Accuracy may deteriorate beyond the model's training ranges or when query rows differ in distribution from context rows.
- NVIDIA has not yet released the artificial data generators or training recipe, saying both will arrive soon.