NVIDIA Kumo Tabular brings open foundation models to tabular prediction

NVIDIA has released Kumo Tabular, an open foundation-model family aimed at a part of machine learning that is often overshadowed by language and vision models: structured tables. The basic idea is in-context prediction. Instead of training a new gradient-boosted model for every dataset, Kumo reads labeled rows as context and predicts labels for new rows in a forward pass. NVIDIA positions the family for both classification and regression without per-task tuning or feature engineering.

The release is unusually compact by current foundation-model standards. NVIDIA describes three sizes ranging from 28 million to 215 million parameters, pretrained only on artificial data. The company provides model weights on Hugging Face and an open-source library, and says the release uses the OpenMDW-1.1 license for commercial use. NVIDIA also reports first-place results across four tabular benchmarks, but those benchmark claims should remain attributed to the publisher until they are independently reproduced under a matching setup.

For the open-weight ecosystem, Kumo is a useful reminder that the category is expanding beyond generative models. A registry focused only on chat LLMs would miss an increasingly important class of downloadable foundation models for speech, robotics, tabular data and other specialized tasks. Open Model Weights can treat Kumo like any other record: verify the exact weight artifacts, preserve the license terms, identify the modality and task correctly, and keep benchmark claims separate from source-verified facts about what is actually downloadable.