Large Tabular Models Excel Where LLMs Fail

2026-07-09 · Source: IEEE Spectrum · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, short

Summary

Fundamental, an AI startup, launched its large tabular model (LTM) NEXUS on 5 February 2026 with US \$275 million in funding, addressing a critical gap where large language models (LLMs) fail: analyzing structured tabular data. While LLMs excel at text and image generation, they struggle with non-sequential spreadsheet data, which is crucial for most organizations. LTMs like NEXUS are foundation models pre-trained on billions of diverse tables, including proprietary, public, and augmented datasets, to directly model tabular data's structure, learning numerical values, representations, and inter-entry relationships for accurate, deterministic predictions. Unlike traditional machine learning algorithms like XGBoost, LTMs require minimal bespoke engineering. Amazon Web Services has already embedded NEXUS in Amazon SageMaker, leveraging its confidential computing platform that prevents access to customer data. Other entities, including Feedzai, Mastercard, Google (with TabFM), and research groups (FlexTab, TabICL, iLTM), are also developing LTMs, signaling a shift towards automated, combined LLM and LTM systems for future data analysis.

Key takeaway

For data scientists and machine learning engineers evaluating solutions for structured data analysis, recognize that Large Tabular Models (LTMs) like Fundamental's NEXUS offer superior, deterministic performance compared to LLMs. Your current XGBoost-based workflows can be streamlined by LTMs' foundational learning, reducing months of bespoke engineering. Prioritize LTMs with confidential computing capabilities, such as NEXUS on Amazon SageMaker, to ensure secure, in-place processing of your sensitive enterprise data without compromising privacy.

Key insights

Large Tabular Models (LTMs) are purpose-built foundation models that excel at analyzing structured data where LLMs fail.

Principles

• Tabular data's non-sequential nature challenges LLM sequence prediction.
• LTMs model joint numerical value, representation, and relational context.
• Deterministic predictions are essential for reliable tabular analysis.

Method

LTMs are pre-trained on billions of diverse tables, including proprietary and augmented datasets, to directly model tabular data structure for broad application.

In practice

• Apply LTMs for fraud detection or clinical trial data analysis.
• Use LTMs for predictive tasks on enterprise spreadsheet data.
• Integrate LTMs for secure, in-place analysis of sensitive datasets.

Topics

• Large Tabular Models
• Tabular Data Analysis
• Foundation Models
• NEXUS
• Confidential Computing
• Amazon SageMaker

Best for: AI Engineer, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Beyond the lakehouse: Fundamental's NEXUS bypasses manual ETL with a native foundation model for tabular data — NEXUS is a Large Tabular Model designed to natively understand and predict from complex, non-sequential business data.
• LLMs on Tabular Data with Limited Semantics: Evidence from Industrial Car Retrofit Prediction — LLMs serve best as complementary components to strong tabular machine learning baselines, not replacements, for industrial data with limited semantics.
• Fundamental’s Large Tabular Model NEXUS is now available on Amazon SageMaker JumpStart — NEXUS is a foundation model purpose-built for tabular data, delivering deterministic, scalable predictions without extensive feature engineering.
• Is XGBoost gone: How Relational Foundation Models Conquered 500 Billion Row Enterprise Data — The dominance of gradient boosted trees on tabular data was an illusion created by pre-processed Kaggle datasets.
• From Rules to Predictions: Federated Tabular Learning with LLM Reasoning — Federated learning can leverage LLM reasoning on privacy-preserving decision rules to improve tabular data model performance without raw data exposure.
• Introducing TabFM: A zero-shot foundation model for tabular data — TabFM applies zero-shot in-context learning to tabular data, eliminating manual tuning and feature engineering for classification and regression.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by IEEE Spectrum.