Databricks has published code comparing five AI models designed to make predictions from spreadsheet-style data, with a practical trade-off: the models may handle large tables, but can cost more per prediction than CPU-based XGBoost. The guide could help businesses assess an alternative to building a separate model for every task, though licensing and memory use deserve a close look.
Databricks Watch analysis
What happened
The models, known as tabular foundation models, are pretrained on synthetic datasets and then use a company’s example rows when making predictions. The Next Web says Databricks’ guide includes free code comparing five models, with uses for this kind of data including credit risk, customer churn and demand forecasting.
According to the report, Databricks says three models, TabICLv2, Mitra v2 and KumoTabular, can be used commercially. TabFM and the newest TabPFN require a separate agreement with their makers. Databricks says its AI Runtime GPU service can run TabPFN-3 on NVIDIA H100s with tables of up to one million rows and 200 columns. The service is in preview and uses pay-as-you-go pricing. Read The Next Web’s report.
Why it matters
Businesses often train and retrain separate models for individual prediction tasks. These pretrained alternatives offer another way to work with tabular data, and the guide gives teams code to compare options rather than asking them to take the idea on faith.
There are costs beyond the GPU bill. The report says memory use rises with the number of rows because the models read the training set at prediction time. Databricks also says they cost more per prediction than XGBoost running on an ordinary processor. So this is not an automatic upgrade: the appeal will depend on a team’s data, workload and licence terms.
Our read
The useful part is the comparison code and the concrete scale Databricks says its service can handle. Teams should test these models against their current approach on their own data, while checking commercial rights and the full cost of inference. “Foundation model” is not a coupon code for cheaper predictions.
What to watch
- Whether independent tests find advantages over established approaches such as XGBoost.
- How memory use and per-prediction costs change with table size.
- Whether the models requiring separate commercial agreements become easier to license.
Discussion spark: For business prediction tasks, would you choose a tabular foundation model for its flexibility, or stick with a cheaper specialised model unless it clearly wins on your own data?
Sources and evidence
- Source update (16 November 2026, 19:28 UTC)
Independent WittyWires tracker for public updates about Databricks. Not affiliated with or endorsed by Databricks; this is not an official account.