NVIDIA has released Structured Data Models, a library that puts several pretrained foundation models for tabular data behind a single PyTorch interface. The goal is to let users make predictions on tables without training a model from scratch. Instead of building and tuning a pipeline, you can load a pretrained model and apply it directly to your data.
The library currently wraps TabICLv2 and KumoTabular for single-table use cases. A tutorial from MarkTechPost demonstrates how to implement the library and run zero-training predictions. The tutorial focuses on practical steps, showing how to access the models and generate outputs with minimal setup.
Because the source is a single tutorial, there are no conflicting accounts to compare. The main takeaway is that NVIDIA is pushing pretrained models beyond text and images into the tabular domain, where traditional machine learning usually requires task-specific training.