Databricks releases guide and code for tabular AI models on GPU service

What happened
On November 16, 2026, Databricks released a technical guide and an open-source repository for running five tabular foundation models on its AI Runtime GPU platform. The release includes tools for comparing different models and testing fine-tuning workflows on enterprise prediction tasks using spreadsheet-like data. These models are pretrained and can make predictions without the need for task-specific training. The solution targets organizations that need to predict outcomes from tabular data such as credit risk or customer churn.
Key details
- The guide and repository cover five models: TabPFN, TabFM, TabICL, Nori, and KumoTabular.
- Some models, like TabFM and TabPFNv3.5, have non-commercial weights and require agreements for commercial use.
- GPU inference for these models costs more per prediction than using XGBoost on a CPU.
- The AI Runtime platform supports on-demand H100 GPUs and can handle up to 1 million rows by 200 features for TabPFN-3.
- The repository provides shared evaluation datasets, metrics, and supports both batch and real-time deployment options.
Why it matters
Organizations using Databricks can now test and deploy tabular foundation models for predictions without building separate training pipelines for each task, potentially saving engineering time. However, they face higher per-prediction costs due to GPU requirements and must consider licensing terms for commercial use. Teams must evaluate the trade-off between reduced development effort and the ongoing expense of GPU-backed inference for their specific data and use cases.

