NVIDIA's Kumo Tabular predicts on spreadsheets with no training

NVIDIA released Kumo Tabular, an open foundation model that makes predictions on tabular data in a single forward pass, with no training, tuning, or feature engineering. You hand it a table and the column you want predicted, and it returns classification or regression results plus an uncertainty estimate, the way a large language model answers a prompt. The models are small, three sizes from 28 million to 215 million parameters, and NVIDIA says the largest runs 17 times faster than LimiX-2 while scoring higher.

What is unusual is the training data: none of it is real. Kumo Tabular was trained entirely on synthetic tables generated from Structural Causal Models, which lets it see a wide range of column relationships without touching any dataset it will later be tested on. The architecture mixes attention across columns, across rows, and in-context, with an attention temperature that scales with table size, and it handles missing values directly instead of filling them in. NVIDIA reports first-place finishes on four benchmarks in its post, including a 1950 ELO on TabArena, the highest overall.

Why it matters

If you build models on spreadsheets or database tables, this is a way to get a strong baseline in seconds without the usual cycle of feature engineering and hyperparameter tuning, worth testing before you spend a week hand-building a gradient-boosted model.