Large Tabular Models Excel Where LLMs Fail
TL;DR
IEEE Spectrum frames Large Tabular Models as a new model class for spreadsheets, because LLMs struggle with structured data: rows and columns are not a linear text sequence. Startup Fundamental launched NEXUS on 5 February 2026 with US $275 million in funding. The model targets tabular prediction and aims to reduce custom feature engineering. AWS embedded NEXUS into Amazon SageMaker in June. Fundamental says NEXUS was pretrained on billions of tables and is not trained on customer data.
Nauti's Take
This is bigger than another model name. LLMs are good at talking about data, but spreadsheet work needs stable predictions, column logic and less hallucination.
That is a real market, because companies do not need another chatbot as much as they need answers from existing data. Still, NEXUS also smells like a polished enterprise story right now: big funding, an AWS badge, big promises.
Real benchmarks and messy customer use cases will decide whether LTMs replace XGBoost or become a new layer on top of classic machine learning.
Briefingshow
Most business data lives in spreadsheets, databases and logs, not chat windows. If LTMs can generalize across tables, teams could build predictions for fraud, marketing, inventory, research or medicine faster. The catch: many claims are vendor-led, and sensitive tabular data remains a serious test for trust, access control and governance.
Sources