NVIDIA has introduced Kumo Tabular, a series of open-source foundation models designed for structured data. The models are available in Small, Medium, and Large configurations, ranging from 28 million to 215 million parameters. They operate within the company’s structured-data-models (SDM) library, which is built for GPU-native preprocessing and inference.
The models are released under the OpenMDW-1.1 license, allowing for commercial use, while the SDM code follows the Apache-2.0 license. Kumo Tabular is trained on synthetic tables generated from Structural Causal Models. The training process involved three stages, with context sizes scaling from 1,024 to 60,000 rows and up to 100 columns.
Performance benchmarks indicate that Kumo Tabular achieves an Elo rating of 1950 on TabArena. NVIDIA reports that the model runs 17 times faster than LimiX-2 on an RTX 6000 Pro GPU. The architecture utilizes column, row, and in-context attention mechanisms, with the head providing class probabilities or 999 quantiles for regression tasks.
The SDM library also supports other models, including TabICLv2, Google’s TabFM, and KumoRelational. All models within the library share a unified interface based on the TableTensor container. NVIDIA plans to release the training recipes and data generators for Kumo Tabular in the near future.
