Large Tabular Models Excel Where LLMs Fail

Ask ChatGPT to write you a sonnet or untangle a contract, and it'll manage fine. Hand it a spreadsheet of transaction logs or clinical trial vitals, and it falls apart. That gap is the whole premise behind Fundamental, a startup that came out of stealth on February 5, 2026 with $275 million in funding and a model called NEXUS, built from the ground up for rows and columns rather than sentences.

The reason large language models stumble here isn't a minor bug — it's baked into how they work. LLMs predict the next token in a sequence, so word order matters enormously; shuffle a sentence and you can destroy its meaning. Tabular data doesn't behave that way. Swap two columns in a spreadsheet or reorder the rows, and the facts underneath don't change at all. Jeremy Fraenkel, Fundamental's CEO, points out that LLMs are also inherently a bit unpredictable — feed one a slightly different prompt and you might get a different answer, which is fine for a chatbot but useless if you're deciding whether a transaction is fraudulent and need the same answer every time.

Until now, the industry's fallback has been gradient-boosted decision trees like XGBoost, tools that have been around for more than 15 years and still require data scientists to hand-tune a new model for every single use case, often over months. NEXUS instead was pretrained on billions of tables, learning not just the numbers in a cell but what they represent and how they relate to everything around them — so it can generalize across tasks instead of starting from scratch each time. Fraenkel says building that training corpus was the hard part, since tabular data is scarce, sensitive, and wildly inconsistent between domains — a hospital's dataset looks nothing like a bank's. Fundamental leaned on proprietary partnerships, licensed data, open-source sets, and augmentation to get there, and Fraenkel is emphatic that NEXUS was never trained on actual customer data; the platform is built so the company physically can't access it.

That design choice likely explains why Amazon Web Services embedded NEXUS into SageMaker in June, plugging the model into what's effectively the default home for secure machine learning at many companies. Instead of importing sensitive data out to a model, the model shows up where the data already lives. Fundamental isn't running unopposed, either — Feedzai and Mastercard rolled out their own finance-focused tabular tools in March, Google followed in late June with TabFM, trained on synthetic data, and academic projects like FlexTab, TabICL, and iLTM are chasing the same idea from the research side.

Boris van Breugel, an Amsterdam-based researcher who co-authored an early paper flagging this blind spot back in 2024, thinks the underlying reason nobody noticed sooner is almost embarrassingly human: people like looking at pictures and chatbot replies, not columns of numbers. Fraenkel frames the eventual endpoint as something like a divided brain, with LLMs handling language and reasoning while LTMs handle statistics and pattern recognition — and the real payoff coming once the two are actually combined.