Large Tabular Models: Why LLMs Struggle with Tabular Data – and What LTMs Do Better

The Blind Spot of Language Models: Structured Data

Large language models – LLMs for short – have demonstrated impressive capabilities in recent years. They draft contracts, solve mathematical problems, generate code, and summarize medical literature. Yet there is one task that appears trivial at first glance where they consistently fall short: the analysis of structured, tabular data.

And yet tables are the backbone of virtually every business. Bank transaction logs, marketing metrics, clinical trial evaluations, inventory lists – all of it exists in rows and columns. Anyone who believes a modern AI chatbot system can reliably interpret this data and derive well-founded predictions from it is mistaken. And this is precisely where a new class of models comes in: the so-called Large Tabular Models (LTMs).

Why LLMs and Tables Are a Poor Match

To understand why LLMs struggle with tabular data, one must consider their underlying architecture. Language models are trained to recognize sequential patterns in text – they predict the next token in a sequence. Language is inherently linear: the order of words fundamentally determines the meaning of a sentence.

Tabular data, by contrast, is non-sequential. Columns can be reordered or rows swapped without altering the factual content of the data. This structural property is fundamentally difficult to reconcile with the Transformer architecture on which all large language models are based. An LLM that receives slightly different inputs produces slightly different outputs – desirable in creative text generation, but dangerous when the question is whether a financial transaction should be classified as fraudulent or not.

Compounding this is the fact that tabular datasets are highly heterogeneous. A biological research dataset and a financial dataset share almost no structural commonalities – unlike natural language, which despite all its variety remains semantically comparable. This makes it extremely difficult to train a single model on a broad base of tabular data.

The Previous Alternative: Gradient-Boosted Decision Trees

Anyone who wanted to apply machine learning to tabular data in the past turned to classical algorithms – above all XGBoost and other gradient-boosted decision tree methods. These technologies have been in use for more than 15 years and are deployed by companies worldwide to build predictive models on structured data.

Their decisive drawback: they must be trained, calibrated, and optimized by data science teams over months for each individual use case. These are highly specialized, labor-intensive processes – without the transferability and scalability that modern foundation models offer.

„Classical algorithms like XGBoost are powerful, but they don't scale like foundation models. Every new use case requires building a new model from scratch – that is personnel-intensive and time-consuming," explains Dr. Maik Bunzel, founder and CEO of mabucon.eu, who supports companies in implementing intelligent AI agents and automation workflows.

What Large Tabular Models Do Differently

The US-based AI startup Fundamental brought a new model category to the forefront in early 2026 with its model NEXUS. NEXUS was developed as a pure foundation model for tabular data – and its approach differs fundamentally from previous solutions.

While LLMs model sequences of tokens, LTMs model the structure of tabular data directly. They simultaneously learn:

• the numerical value of an entry
• what that value represents in terms of content
• how the entry relates to other columns and rows in the table
• the statistical properties of the entire data distribution

This contextual understanding enables more precise inferences and predictions. An LTM processing a warehouse inventory entry for bananas does not merely understand the number 500 – it understands that this is a quantity figure belonging to the product category "fresh produce," and how it relates statistically to other entries in the table.

Also crucial is the determinism property: Unlike LLMs, which can produce different outputs for slightly altered inputs, LTMs are designed to deliver stable and reproducible predictions – an indispensable characteristic for business-critical decisions such as credit approval, fraud detection, or quality assurance.

Training on Billions of Tables – and the Data Problem

One of the greatest challenges in building LTMs lies in sourcing training data. Natural language is available in enormous quantities on the internet and is structurally relatively homogeneous. Tabular data, by contrast, is often sensitive, proprietary, and structurally extremely diverse.

According to its own statements, Fundamental pre-trained NEXUS on billions of tables – using a combination of licensed datasets, public sources, and purpose-built data augmentation techniques. The company explicitly emphasizes that customer data is neither used for training nor can it be accessed by Fundamental. NEXUS operates as a Confidential Computing Platform, which is particularly relevant from a data protection standpoint.

This privacy architecture is likely to have been a key factor in Amazon Web Services (AWS) integrating NEXUS into Amazon SageMaker in June 2026 – one of the most widely used platforms for secure enterprise machine learning. This integration now makes LTM technology accessible to a broad range of enterprises.

Implications for Businesses: What LTMs Mean in Practice

The emergence of powerful Large Tabular Models has far-reaching consequences for companies that rely on data-driven decisions. Until now, data science teams had to develop and maintain a separate model for each new use case. With foundation models for tabular data, this underlying logic changes: a pre-trained model can be transferred to various prediction tasks with minimal adaptation.

For AI-powered automation workflows, this represents a quantum leap. Dr. Maik Bunzel, founder and CEO of mabucon.eu, sees a direct relevance to the automation of complex business processes in this development: "Many of our clients are sitting on enormous amounts of structured data – sales figures, inventory levels, customer transactions. If LTMs deliver on their promise, AI agents could finally truly understand this data and respond to it autonomously."

Concrete application areas that open up include:

• Automated anomaly detection in financial transactions without manual model tuning
• Prediction-driven inventory planning in real time, directly based on ERP data
• Dynamic risk scoring in the lending and insurance sectors
• Quality control in production based on sensor-generated measurement data
• Predictive Maintenance, i.e. anticipatory servicing, without elaborate feature engineering phases

Outlook: A New Era for Structured Data

The emergence of Large Tabular Models marks an important step in the maturation of the AI landscape. It is no coincidence that this area has long been overlooked: structured data is less spectacular than generated images or chatbot responses, but it is the operational backbone of virtually every business.

With the advent of powerful, scalable LTMs, a technological gap is being closed that is critical for the practical deployment of AI in enterprises. Integration into established platforms such as AWS SageMaker significantly lowers the barrier to entry. At the same time, further providers will develop their own LTM approaches – a competition that is likely to accelerate the technology's maturation rapidly.

For companies seriously considering the use of autonomous AI agents, developments in this area deserve close attention. Because only when AI understands not just text, but also the structured data on which business processes are built, can it truly execute those processes autonomously and reliably – and thereby unlock its full automation potential.