NVIDIA releases open Kumo Tabular foundation model for zero-shot tabular prediction

NVIDIA has released Kumo Tabular, an open foundation model for tabular classification and regression that operates via in-context learning without requiring training or feature engineering. Pretrained exclusively on artificial data generated by Structural Causal Models, the model is available in three sizes ranging from 28M to 215M parameters.

• Kumo Tabular ranks first on the TabArena, BeyondArena, TALENT, and ScoringBench benchmarks, establishing a new state of the art on the accuracy-efficiency Pareto front.
• The architecture uses column, row, and in-context attention to process tables, allowing predictions for new rows in a single forward pass.
• It runs via NVIDIA's open-source structured-data-models library under the OpenMDW-1.1 license for commercial use.

The release provides enterprises with a zero-shot alternative to gradient-boosted trees, enabling immediate prediction on labeled tables without the traditional lifecycle of hyperparameter tuning and validation.