NVIDIA Kumo Tabular Sets New Accuracy-Efficiency Frontier for Tabular Prediction

NVIDIA's Kumo Tabular, an open foundation model for tabular data, has achieved an overall ELO score of 1950, ranking first in multiple benchmarks including TabArena, BeyondArena, TALENT, and ScoringBench. The model is available on Hugging Face and comes in sizes ranging from 28M to 215M parameters.

Kumo Tabular was trained on approximately 35 million, 71 million, and 137 million artificial tables for its Small, Medium, and Large variants, respectively. Its performance is notable for running 17 times faster than the previous LimiX-2 model, significantly enhancing efficiency in tabular data predictions.

In the TALENT benchmark, Kumo Tabular achieved average ranks of 6.67 for classification accuracy, 3.98 for classification log-loss, and 4.22 for regression RMSE. The model also boasts an Improvability score of 7.78%, indicating its potential for further enhancements.

Kumo Tabular operates effectively on numerical and categorical columns, returning class probabilities or numeric predictions in a single forward pass. However, NVIDIA cautions that accuracy may degrade on tables that exceed the training ranges, which span from 1,024 rows to a maximum of 60,000 rows across three training stages.

The evaluation of Kumo Tabular utilized the RTX 6000 Pro GPU, underscoring the hardware requirements for optimal performance. This development follows NVIDIA's recent initiatives in AI agent security, highlighting the company's ongoing commitment to advancing AI capabilities across various domains.