Trending
Prev Story
Next Story

Databricks pitches AI that predicts from tables without training

The Next Web (TNW) •
November 16, 2026 • 07:28 PM
• By Cristian Dina
3
3 reads 0 reactions

Databricks has published a guide and free code for tabular foundation models, AI that predicts from business data without task training. GPUs cost more.

1 / 5
Swipe Photos

Flapping Airplanes wants to train AI with less data and get more out of each chip.

Databricks has published a guide and open code for running five tabular foundation models, AI that makes predictions from spreadsheet-style data without task-specific training, on its AI Runtime GPU service. It says the models cost more per prediction than XGBoost on CPUs, and two need a commercial deal with their makers.

Databricks has published a guide to running “ tabular foundation models ” on its platform, the company said in a blog post on Monday. These AI models make predictions from spreadsheet-style data without being trained for each task. The guide comes with free code that compares five of them.

Businesses use table data to predict things like credit risk, customer churn and demand. Usually, teams build and retrain a separate model for every task. Tabular foundation models are instead pretrained on millions of made-up datasets, then learn from a company’s example rows at the moment they predict.

The idea is catching on. SAP, which works closely with Databricks , closed a €1B-plus deal for Prior Labs, maker of the TabPFN model, in July. H2O.ai launched its own model in May.

Databricks is open about the trade-offs. Running these models on GPUs costs more per prediction than a classic XGBoost model on an ordinary processor, it says. Memory use also grows with every row, because the model reads the whole training set each time.

Licences are another catch. Of the models Databricks lists, TabICLv2, Mitra v2 and KumoTabular can be used commercially. TabFM and the newest TabPFN need a separate deal with their makers.

The pitch is for Databricks’ AI Runtime, a pay-as-you-go GPU service in preview since March. Databricks says its Nvidia H100 chips can run TabPFN-3 on tables of up to 1 million rows and 200 columns. The company, recently valued at $190B , still advises firms to test the models on their own data first.

Get the most important tech news in your inbox each week.

Prev Swipe Story Next

Related Stories

Technology
Build #2612