Large language models (LLMs) have revolutionized artificial intelligence by generating text, images, and code, but they remain surprisingly poor at interpreting structured data stored in spreadsheets and databases. An IEEE Spectrum article highlights the emergence of large tabular models (LTMs), a new category of foundation models designed specifically for tabular data. Leading this effort is NEXUS, developed by AI startup Fundamental, which aims to give businesses a faster and more accurate way to analyze structured information.
The article explains that enterprise data is fundamentally different from natural language. Financial records, medical data, scientific measurements, and customer transactions are organized into rows and columns, where the order of entries often has little impact on their meaning. LLMs, however, are built to process sequential information by predicting the next token in a sequence. As a result, they struggle with spreadsheets because rearranging columns or rows changes the input without changing the underlying data. NEXUS overcomes this limitation by modeling relationships within tables directly rather than treating them as text.
Unlike traditional machine learning systems, which often require months of feature engineering and model tuning for each application, NEXUS is pre-trained on billions of tabular datasets collected through partnerships, licensed sources, public datasets, and data augmentation. It learns from numerical values, categories, and the relationships among entries, allowing it to perform prediction, classification, and anomaly detection across different industries with minimal customization. Fundamental also emphasizes that customer data remains private because the platform is designed so the company cannot access or train on users’ proprietary information.
The technology has already gained commercial traction through an integration with Amazon SageMaker, enabling organizations to apply the model to sensitive enterprise datasets within their own cloud environments. Although competitors are developing similar systems, the rise of large tabular models signals a broader shift in AI. Rather than forcing every task through language models, the next generation of foundation models will be tailored to specific types of data, allowing businesses to unlock insights from the structured information that drives finance, healthcare, manufacturing, logistics, and scientific research.
