---
title: "Tabular foundation models on Databricks AI Runtime"
description: "Tabular foundation models (TFMs) are pretrained transformers that make predictions on new tabular datasets without task-specific training, using in-context learning on labeled rows passed as context. They eliminate the need for bespoke model development and maintenance pipelines by inferring relationships from millions of diverse synthetic datasets used during pretraining. However, TFMs require GPU infrastructure because their transformer architecture and in-context learning demand dense matrix multiplications and large memory for activations, which grows with dataset size. Databricks AI Runtime addresses this by providing managed serverless GPU infrastructure directly where data resides, with integrated observability and governance through Unity Catalog. The platform supports zero-shot inference and fine-tuning of models like KumoTabular, enabling enterprises to deploy TFMs as batch pipelines or real-time endpoints without creating data silos."
---

# Tabular foundation models on Databricks AI Runtime

[Databricks](https://yomu.fyi/company/databricks) · Marc Susagna Holgado, Javier Poveda Panter, Puneet Jain, Ryuta Yoshimatsu · Nov 16, 2026

**Type:** Explainer

## Summary

Tabular foundation models (TFMs) are pretrained transformers that make predictions on new tabular datasets without task-specific training, using in-context learning on labeled rows passed as context. They eliminate the need for bespoke model development and maintenance pipelines by inferring relationships from millions of diverse synthetic datasets used during pretraining. However, TFMs require GPU infrastructure because their transformer architecture and in-context learning demand dense matrix multiplications and large memory for activations, which grows with dataset size. Databricks AI Runtime addresses this by providing managed serverless GPU infrastructure directly where data resides, with integrated observability and governance through Unity Catalog. The platform supports zero-shot inference and fine-tuning of models like KumoTabular, enabling enterprises to deploy TFMs as batch pipelines or real-time endpoints without creating data silos.

## Context

Enterprises face high costs and complexity in developing and maintaining bespoke machine learning models for tabular data tasks. Tabular foundation models (TFMs) offer a pretrained alternative that eliminates task-specific training, but they require GPU infrastructure that is difficult to provision and manage. This post explains how TFMs work, compares leading models, and describes how Databricks AI Runtime provides managed GPU infrastructure, observability, and governance to support TFM deployment.

## Approach / What changed

Databricks AI Runtime provides managed serverless GPU infrastructure directly where data resides, with integrated observability through Genie and governance through Unity Catalog. It supports zero-shot inference and fine-tuning of TFMs, enabling deployment as batch pipelines or real-time endpoints without moving data out of the platform.

## Takeaways

- TFMs use in-context learning: labeled rows are passed as context alongside rows to predict, and attention relates new rows to those examples, producing predictions in a single forward pass without training.
- TFMs require GPUs because their transformer architecture and in-context learning demand dense matrix multiplications and large memory for activations, which grows with dataset size.
- Databricks AI Runtime eliminates the need to copy data out of the data platform by providing managed GPU infrastructure directly where data resides, with governance through Unity Catalog.

## FAQ

### What are Tabular Foundation Models?

TFMs are pretrained transformer models designed for tabular data. They are pretrained on millions of diverse synthetic datasets so they can infer relationships in new data without needing to be trained from scratch, drastically reducing the time to deploy machine learning models.

### Why are GPUs recommended for TFMs?

TFMs are built on the transformer architecture, which relies on dense matrix multiplications that run vastly more efficiently on GPUs than CPUs. While you can run a TFM on a CPU, it is impractically slow for enterprise-scale data; the massive intermediate calculations during in-context learning require a GPU's parallel processing power and memory bandwidth.

### How does Databricks AI Runtime help with TFMs?

Databricks AI Runtime provides fully managed, serverless GPU infrastructure directly where your data lives. It eliminates the need to move data into external GPU tools, allowing you to run inference, fine-tune models, and manage governance through Unity Catalog without creating data silos.

**Tags:** [Databricks AI Runtime](https://yomu.fyi/topic/databricks-ai-runtime), [GPU](https://yomu.fyi/topic/gpu), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Tabular Foundation Models](https://yomu.fyi/topic/tabular-foundation-models)

- Source: [Databricks](https://www.databricks.com/blog/tabular-foundation-models-databricks-ai-runtime)
- Source URL: https://www.databricks.com/blog/tabular-foundation-models-databricks-ai-runtime
- Ingested by Yomu: 2026-11-16T18:00:12.856Z

[Read original post](https://www.databricks.com/blog/tabular-foundation-models-databricks-ai-runtime)
