How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models

NVIDIA has released Structured Data Models, a library that consolidates several pretrained foundation models for tabular data behind a single PyTorch interface. The library supports models including TabICLv2 and KumoTabular, allowing practitioners to make predictions on structured datasets without any training step.

The zero-training approach is notable because it removes the need to collect labeled data or tune hyperparameters for each new task. Users can load a pretrained model and apply it directly to their tabular inputs, which could significantly reduce the time and expertise required to deploy tabular ML in production.

This development extends the foundation-model paradigm—previously dominant in language and vision—into the structured data domain. Tabular data remains a major use case in industries like finance, healthcare, and operations, where traditional deep learning has historically underperformed compared to gradient-boosted trees. By offering pretrained models that generalize across tables, NVIDIA aims to close that gap.

The unified interface also lowers the barrier for teams that may not have deep expertise in model selection or fine-tuning. A single API to access multiple pretrained architectures simplifies experimentation and integration into existing PyTorch pipelines.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.