Nvidia releases an open model that predicts directly from tables

Nvidia has released Kumo Tabular, an open foundation model that can make predictions from a labeled table in a single forward pass. It handles classification and regression without training a separate model, tuning hyperparameters or manually engineering features for each task.

The model was pretrained entirely on artificial data and comes in 28-million, 88-million and 215-million-parameter versions. Nvidia says it ranked first on TabArena, BeyondArena, TALENT and ScoringBench. Code and weights are available through an open-source library and Hugging Face under the OpenMDW 1.1 license, which permits commercial use.

Kumo brings in-context learning—the ability to infer a task from examples—to records such as transactions, claims and sensor logs. It does not remove the need to validate predictions on a company’s own data, especially for high-stakes decisions. Its practical appeal is a shorter path from a labeled table to a baseline prediction.