NVIDIA launches Kumo Tabular, a tabular foundation model that predicts new rows in one forward pass

NVIDIA has introduced Kumo Tabular, a family of tabular foundation models built for classification and regression tasks. The company said the models use labeled rows as context and can predict a new row in a single forward pass, without training, hyperparameter tuning, or feature engineering. The lineup includes small, medium, and large versions, with parameter counts ranging from about 28 million to 215 million. Kumo Tabular runs through NVIDIA’s open-source Structured Data Models, or SDM, library, and its weights are released under the OpenMDW-1.1 license, which allows commercial use. NVIDIA said the models were fully pre-trained on synthetic tables sampled from structural causal models, while also incorporating real-world data patterns such as missing values and high-cardinality categories. According to the NVIDIA team, Kumo Tabular ranked first on the TabArena benchmark with an Elo score of 1950 under default settings. The team also reported that it runs 17 times faster than LimiX-2 on a single RTX 6000 Pro. The item was cited by MarkTechPost in the Techub News report.