NVIDIA has released Kumo Tabular, a new family of tabular foundation models that predict new rows in a single forward pass without requiring training or hyperparameter tuning. The models are available in Small, Medium, and Large sizes, ranging from approximately 28M to 215M parameters.

  • Weights are licensed under OpenMDW-1.1, permitting commercial use, unlike competitors such as TabPFN-3 and LimiX-2.
  • The models run via NVIDIA's open-source structured-data-models (SDM) library on GPU hardware.
  • Pretraining utilized synthetic tables generated from Structural Causal Models to handle messy real-world patterns like missing values.
  • NVIDIA reports first place on TabArena with an Elo of 1950, and the Large model ranks first on ScoringBench.
  • Inference is reported to be 17x faster than LimiX-2 on a single RTX 6000 Pro while maintaining accuracy.

The release provides a commercially viable, high-performance alternative for tabular data tasks that eliminates the need for traditional feature engineering and model training.