Understand this story
On November 16, 2026, Databricks released a technical guide explaining how to run tabular foundation models (TFMs) on AI Runtime. The guide explains that these pretrained models can make predictions on new datasets without task-specific training, eliminating the need for custom training and retraining workflows. A subsequent Databricks Blog post explained that AI Runtime provides managed GPUs, MLflow tracking, and Unity Catalog integration, supports TFM experimentation and fine-tuning, and compared models including TabPFNv2, TabICLv1, TabPFNv3, TabFM, and KumoTabular in terms of accuracy, inference latency, GPU memory, and cost. Another report stated that the guide is accompanied by a public repository for comparing five models—TabPFN, TabFM, TabICL, Nori, and KumoTabular—and for testing fine-tuning workflows, and noted that during zero-shot inference, GPU memory grows with the number of rows and features, the cost per prediction is higher than that of XGBoost on CPU, and what is mainly saved is engineering and maintenance time.
AI-generated from public reports
Developments
1 confirmed developments
Databricks releases technical guide for tabular foundation models on AI Runtime
Source:Databricks Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 17
Unite.AI
Databricks Releases Guide to Running Tabular Foundation Models, Comparing Five Models and Fine-Tuning Workflows
On November 16, 2026, Databricks released a technical guide explaining how to run tabular foundation models on AI Runtime, along with a public repository for comparing five models—TabPFN, TabFM, TabICL, Nori, and KumoTabular—and testing fine-tuning workflows. The guide notes that these models use labeled rows as context during zero-shot inference, predicting without updating parameters, but GPU memory grows with the number of rows and features, and the cost per prediction is higher than XGBoost on CPU; the main savings come from engineering and maintenance time. Regarding licenses, TabICLv2, Mitra v2, and KumoTabular are commercially usable, while the weights of TabFM and TabPFNv3.5 are non-commercial.
Databricks BlogFirst-party
Tabular Foundation Models on Databricks AI Runtime
Databricks introduced tabular foundation models (TFMs): these pretrained models can make predictions on new datasets without task-specific training, eliminating the need for custom training and retraining workflows. The article compares models such as TabPFNv2, TabICLv1, TabPFNv3, TabFM, and KumoTabular in terms of accuracy, inference latency, GPU memory, and cost, and explains that Databricks AI Runtime provides managed GPUs, MLflow tracking, and Unity Catalog integration to support TFM experimentation and fine-tuning.