NVIDIA Releases Kumo Tabular: Enabling Foundation Models to Be Used Directly on Tabular Data
NVIDIA Releases Kumo Tabular: Enabling Foundation Models to Be Used Directly on Tabular Data
NVIDIA has open-sourced Kumo Tabular on Hugging Face, an open foundation model for tabular classification and regression tasks. Given a labeled sample table and data to be predicted, the model can directly output results through a single forward inference pass, without the need to retrain, tune parameters, or perform feature engineering for each task.
Kumo Tabular is available in three versions—Small, Medium, and Large—with parameter counts ranging from 28 million to 215 million. The model is pretrained entirely on synthetic tabular data and understands table structure through column, row, and context attention; the official release also includes accompanying code and model weights, under the OpenMDW-1.1 license, and can be used for commercial scenarios.
Official tests show that Kumo Tabular has achieved leading results on benchmarks such as TabArena, BeyondArena, TALENT, and ScoringBench. However, it is currently mainly oriented toward numerical and categorical columns, and for data with pronounced distribution shifts, teams still need to use their own held-out data to validate accuracy and calibration performance.
The Encore Aria editorial team believes that the value of Kumo Tabular lies not only in benchmark rankings, but more in how it brings the way foundation models are used to the structured data commonly found in enterprises, such as customers, transactions, sensors, and orders. For teams that previously relied on gradient-boosted trees, feature engineering, and per-task training, this "provide examples and predict" approach is worth continued attention.