Large Tabular Models Outperform LLMs on Structured Data
Google/DeepMind
Feedzai
Mastercard
Amazon Web Services
A new type of generative AI, the large tabular model (LTM), is emerging to handle structured data like spreadsheets, which large language models (LLMs) struggle with. Startup Fundamental has released NEXUS, an LTM adopted by AWS, while Google and others have launched competitors. LTMs model tabular data directly, enabling accurate predictions without the sequential limitations of LLMs.
Large language models (LLMs) behind chatbots like ChatGPT, Claude, and Gemini excel at text and images but fail at analyzing structured data like spreadsheets, which are crucial for banks, marketing agencies, and scientific research. Startup Fundamental emerged from stealth on 5 February 2026 with $275 million in funding and launched NEXUS, a large tabular model (LTM) purpose-built for tabular data. Unlike LLMs, which model sequences, LTMs jointly learn numerical values, their meanings, and relationships, providing deterministic predictions independent of column order. NEXUS was pre-trained on billions of tables via partnerships, public datasets, and augmentation, but not on customer data—it runs on a confidential computing platform. In June, Amazon Web Services embedded NEXUS in Amazon SageMaker. Competitors include Feedzai, Mastercard, and Google with TabFM. Researchers have also released FlexTab, TabICL, and iLTM. Experts predict automated data analysis will dominate, with LTMs and LLMs complementing each other like the human brain's hemispheres.
- Сокращения
- LLM = Large Language Model — большая языковая модель
- LTM = Large Tabular Model — большая табличная модель
- AWS = Amazon Web Services — Amazon Web Services
Source: IEEE Spectrum AI —
original