Fundamental NEXUS: Large tabular model for spreadsheets

Fundamental has built a foundation model called NEXUS that tackles spreadsheets and tabular data, described in the piece as "AI’s surprising final frontier." The article frames NEXUS as a dedicated large tabular model intended to perform tasks that conventional large language models struggle with when faced with structured tables.

What is NEXUS?

NEXUS is a foundation model created by the startup Fundamental specifically to work with tabular data and spreadsheets. The company’s team shown in the piece includes CEO Jeremy Fraenkel, Chief Science Officer Marta Garnelo, and cofounder Gabriel Suissa, who are identified as the founders behind the effort to build an AI model for tables.

The coverage presents NEXUS as a purpose-built model rather than a generic conversational LLM repurposed for tables. The reporting links the model to the broader idea of "large tabular models," implying a class of foundation models trained or adapted for structured numeric and categorical data rather than free-form text.

How does a large tabular model differ from an LLM?

Large tabular models are designed to operate on rows, columns, and the implicit schema of spreadsheets, whereas LLMs are optimized for sequences of text and general language tasks. The headline and subhead state the central claim: large tabular models excel where LLMs fail, positioning tabular work as a distinct technical challenge from natural-language tasks.

The article highlights the distinction by naming the problem domain—spreadsheets and tabular data—as one that NEXUS aims to address. That framing suggests a structural mismatch between general LLM capabilities and the demands of table-oriented problems, such as preserving schema, handling numeric aggregation, and following cell-level logic.

Why does this matter?

If large tabular models can reliably outperform LLMs on spreadsheet tasks, organizations that depend on table-based workflows stand to gain more accurate automation and analysis. Many business processes, scientific datasets, and operational systems rely on structured tables; a model that understands that structure could reduce the need for manual cleaning, mapping, and ad hoc prompts.

The shift also implies a specialization trend in foundation models: instead of one-size-fits-all LLMs, teams may build verticalized models tailored to data modalities. That can change procurement and deployment choices for enterprises that need precise handling of numeric and categorical data rather than conversational fluency.

What to watch

Look for published benchmarks or demonstrations comparing NEXUS to mainstream LLMs on concrete spreadsheet tasks, and for adoption signals from enterprises that manage large volumes of tabular data. Another clear milestone will be practical integrations showing NEXUS applied inside existing spreadsheet tools or data pipelines.

The founders named in the piece—Jeremy Fraenkel (CEO), Marta Garnelo (Chief Science Officer), and Gabriel Suissa (cofounder)—are the immediate points of contact for follow-ups or technical briefings on NEXUS and how Fundamental intends to measure tabular model performance.

Fundamental’s NEXUS puts tabular data squarely in the spotlight. The article presents the model as an explicit attempt to solve spreadsheet-specific failures of LLMs, and the coming months should reveal whether tabular-focused foundation models deliver clear, measurable advantages over general-purpose language models.