Large Tabular Models Outperform LLMs on Structured Data

Large language models (LLMs) that power generative chatbots struggle with analyzing structured tabular data — the primary data format for business. The startup Fundamental has developed a new AI model — a large table model (LTM) called NEXUS — designed specifically for working with tables. Unlike LLMs, NEXUS models the structure of tabular data directly, processing not only numerical values but also the context of their meaning and relationships. The model is pretrained on billions of tables and does not use client data. Amazon Web Services has integrated NEXUS into Amazon SageMaker. Competitors are also emerging: Google launched the TabFM model, while Feedzai and Mastercard released their own LTMs for the financial sector. Researchers have presented models such as FlexTab, TabICL, and iLTM. According to experts, the future lies in automated data analysis, which can be performed by combinations of LLMs and LTMs.