Source: www.databricks.com/blog/tabular-foundation-models-databricks-ai-runtime
Tabular foundation models (TFMs) are pretrained on large numbers of synthetic datasets and predict on new tables by passing labeled rows in as context, a method called in-context learning. Databricks says this removes the need for bespoke training and retraining pipelines for each prediction task, although teams can also fine-tune a TFM's weights on their own data.
The main obstacle is GPU use, since in-context learning passes the full training set with every prediction and memory grows with row and feature counts. Databricks says AI Runtime offers on-demand GPUs, including H100s, for TFMs such as TabPFN-3 at up to 1M rows by 200 features. It also provides MLflow experiment tracking, Unity Catalog governance, GPU usage observability, and real-time REST endpoints through Databricks Model Serving.
The post compares several models by licensing. TabICLv2 is under the BSD 3-Clause license, Mitra v2 is Apache-2.0, and KumoTabular is under OpenMDW-1.1, all described as commercially permissive. TabFM and TabPFNv3.5 are not commercially permissive, and the post recommends testing any model directly on an organization's own data.
Why it matters if you build agents
Teams building agents that need tabular predictions can test a pretrained model before building a training pipeline, but should budget for GPU inference, which the post says costs more per prediction than a traditional model such as XGBoost on a CPU. Check each model's license before shipping, because TabFM and TabPFNv3.5 are not commercially permissive.
Get stories like this every weekday
One short email with the day's agent tooling news.