- Published on
NVIDIA Releases Kumo Tabular: Open Tabular Foundation Models That Predict New Rows in a Single Forward Pass
- Authors
- Name
- DeepHorus
NVIDIA Releases Kumo Tabular Open Foundation Models
NVIDIA has released Kumo Tabular, foundation models for classification and regression. They predict new rows in one forward pass from labeled examples, without task-specific training, tuning or feature engineering. The approach could simplify structured-data prediction.
Small, Medium and Large versions span 28 million to 215 million parameters. Kumo combines column, row and in-context attention in a Transformer, following TabICL and TabPFN. Type-specific Fourier features encode cells, while four [CLS] tokens compress rows.
Query rows attend only to labeled context rows, allowing cached context values to be reused. A learned temperature based on the logarithm of key count keeps attention sharp on larger tables. The head returns class probabilities or 999 regression quantiles, adding uncertainty estimates.
Kumo runs through NVIDIA’s open-source structured-data-models library and requires Python 3.11+ and PyTorch 2.7+. Weights carry the commercial OpenMDW-1.1 license, while library code is Apache-2.0. Its TableTensor interface also supports preprocessing, ensembling, TabICLv2, Google’s TabFM and KumoRelational.
NVIDIA trained Kumo entirely on synthetic structural-causal-model tables. Context expanded from 1,024 to 60,000 rows and up to 100 columns; generators added missing values, high-cardinality categories and heavy-tailed targets. Small, Medium and Large saw roughly 35 million, 71 million and 137 million tables.
In comparisons checked September 30, 2026, NVIDIA reports first overall on TabArena at 1950 Elo and inference 17 times faster than LimiX-2 on one RTX 6000 Pro. All three sizes reached the accuracy-speed Pareto front. Kumo also led BeyondArena at 1418 Elo, scored 7.78% Improvability and ranked strongly on TALENT.
Commercial licensing distinguishes Kumo and TabICLv2 from the non-commercial weights of TabPFN-3, LimiX-2 and TabFM. NVIDIA’s benchmark claims and synthetic-only training leave questions about real-world performance. The company plans to release its recipe and data generators, enabling closer reproduction and evaluation.
