Large-language models have been the foundation for the AI boom, helping companies write code and analyze reams of corporate data that no one considered valuable until their debut. But a ton of corporate data sits in structured, tabular formats like spreadsheets that LLMs are less suited for parsing.
Tabular foundation models (TFM) were already on Google and IBM's radar as a possible solution to that problem, but SAP’s announcement last month that it was acquiring TFM startup Prior Labs and investing more than €1 billion ($1.16 billion) into it brought a fresh spotlight to the category.
TFMs first emerged around the release of TabPFN in 2023, and Prior Labs launched in 2024 to oversee and commercialise that model. TabPFN was trained on tabular datasets to address the problems that LLMs face trying to process data in an unfamiliar form and make predictions about future outcomes for enterprise users.
“If you're using unstructured text, the problem has been solved. There are enough LLMs out there. When you talk about tabular data, [that] problem has not been solved,” Giuliano Pezzoli, CEO of Schema Labs, told The Stack.
In the years since TabPFN’s release, more researchers invested in the idea; IBM released its TabularFM framework in late 2024 and Google unveiled its “zero shot” model TabFM in June 2026, touting its ability to process an entire dataset as a “single unified prompt.”
But despite the attention from some of tech’s biggest companies, TFMs remain a fairly niche part of the AI conversation. Wider understanding of how they work and what they can do, however, could make them a major part of enterprise AI sooner rather than later.
