NVIDIA has released Kumo Tabular, a family of open-weights tabular foundation models ranging from 28 million to 215 million parameters for classification and regression tasks. Operating via NVIDIA’s open-source structured-data-models (SDM) Python library, the model uses in-context learning to predict output rows in a single forward pass without requiring feature engineering or hyperparameter tuning.
The model weights are released under the OpenMDW-1.1 license, which permits commercial use, while the underlying SDM library is distributed under Apache-2.0. Kumo Tabular was pretrained entirely on synthetic tables generated from Structural Causal Models that incorporate realistic data patterns, such as missing values, high-cardinality categories, and heavy-tailed target distributions.
According to NVIDIA, Kumo Tabular achieved first place on the TabArena benchmark with an Elo score of 1950, while executing 17 times faster than LimiX-2 on a single RTX 6000 Pro GPU. The launch provides developers with a commercially permissive alternative to existing proprietary or non-commercial tabular models.
Why it matters
Commercially permissive foundation models bring single-pass, zero-shot predictions to enterprise tabular datasets.
Synthetic data pretraining via Structural Causal Models offers a practical blueprint for domain-specific model training.
GPU-native structured data libraries significantly reduce inference latency compared to traditional tabular benchmarks.
Source: marktechpost.com



