New Large Tabular Models Aim to Revolutionize Structured Data Analysis

While current large language models (LLMs) like ChatGPT excel at generating human-like text and images, they falter when analyzing structured data, a critical component for most organizations. This limitation stems from LLMs' sequential processing, which is incompatible with the non-linear nature of tabular data found in spreadsheets. AI startup Fundamental is addressing this gap with its new Large Tabular Model (LTM) called NEXUS, which was launched on February 5, 2026, with $275 million in funding. NEXUS is specifically designed to understand and process row-and-column data, unlike LLMs that are trained on vast amounts of text.

Boris van Breugel, a senior AI researcher, notes a human bias towards visually appealing data, leading to less attention on tabular formats. He also points out that while language has consistent semantics, tabular datasets vary greatly in their variables, making a single model for all tables challenging. Jeremy Fraenkel, CEO of Fundamental, emphasizes that unlike LLMs, which can produce varied outputs, LTMs need to be deterministic for critical predictions like fraud detection. Traditional methods for tabular data analysis, such as gradient-boosted decision trees, require extensive manual optimization by data scientists over months for each specific use case.

Fundamental's NEXUS, however, is a foundational model pre-trained on diverse tabular datasets, allowing for broader application with minimal customization. It directly models the structure of tabular data, understanding the context and relationships between entries. Amazon Web Services has integrated NEXUS into Amazon SageMaker, enabling secure access to customer data without it needing to be imported to the model. Other companies like Feedzai and Mastercard, along with tech giants like Google with its TabFM, are also developing similar LTMs, indicating a significant industry shift towards specialized models for structured data analysis.