NVIDIA releases Kumo Tabular for classification and regression

NVIDIA has released Kumo Tabular, a pretrained foundation model built for classification and regression on structured data. Downloadable weights and an inference package give developers a reusable model for predicting targets on new rows from labeled examples provided as context.

The model card points developers to NVIDIA’s structured-data-models package and demonstrates Kumo Tabular running a classification task on CUDA. NVIDIA calls the library GPU-native and lists Python 3.11 or later and PyTorch 2.7 or later as its baseline. That developer surface differs from NVIDIA’s PAIR beta for local AI inference, which routes requests among compatible home PCs.

Kumo Tabular’s API accepts numerical inputs and categorical or numerical targets for classification or regression. The documentation does not list support for multiple targets or related tables.

NVIDIA says the model first converts table cells into fixed-size row representations, then applies a dataset-wide in-context learning transformer. In that second stage, labeled context rows are used to predict targets for query rows, allowing inference without updates to the pretrained weights.

NVIDIA says Kumo Tabular establishes a new accuracy-efficiency frontier. The reviewed evidence does not independently establish that benchmark-leadership claim or show downstream production adoption.

NVIDIA offers small, medium, and large variants. Its model overview lists parameter counts ranging from 27.46 million for small classification to 215.68 million for large regression. NVIDIA says the Kumo Tabular weights are available under the OpenMDW 1.1 license.