NVIDIA has released Kumo Tabular, an open foundation model for tabular classification and regression, now available on Hugging Face as part of the NVIDIA Kumo Structured model collection. The model is designed to predict labels for new rows in a single forward pass using in-context learning, allowing for predictions without manual training, tuning, or feature engineering.
Model ReleasesNvidiaKumo Tabular
NVIDIA Releases Kumo Tabular: An Open Foundation Model for Efficient Tabular Prediction
Kumo Tabular is built on a Transformer architecture specifically optimized for tabular structures, utilizing column, row, and in-context attention mechanisms. It can handle numerical and categorical data, with built-in preprocessing capabilities to manage text, images, or timestamps. To maintain precision as table sizes increase, the model implements "Length-aware Attention Temperature," which scales attention based on the logarithm of the number of keys.
The model was pretrained entirely on artificial data generated through Structural Causal Models (SCM), which produced millions of diverse and imperfect tables to ensure robustness. Kumo Tabular is available in three sizes, with parameter counts ranging from 28M to 215M. In evaluations across several benchmarks, including TabArena, BeyondArena, TALENT, and ScoringBench, Kumo Tabular achieved top rankings, establishing a new state-of-the-art on the accuracy-efficiency Pareto front.
The model is released under the OpenMDW-1.1 license for commercial use and is optimized to run via NVIDIA's GPU-native library for structured-data-models.
Sources
- NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction （Hugging Face Blog, 2026-09-29）
- open-source library