AWS has announced official support for Fundamental's NEXUS model on Amazon SageMaker AI. With this launch, enterprises can deploy a foundation model (FM) specifically engineered for tabular data prediction. NEXUS helps organizations generate accurate, deterministic predictions from structured datasets in days rather than months. In this post, we explore how to get started with NEXUS on Amazon SageMaker JumpStart, review the deployment process, and demonstrate running predictions against enterprise data.
What is NEXUS? It is a Large Tabular Model (LTM) developed by Fundamental, custom-built for tabular prediction. While LLMs excel at processing text and traditional ML requires tedious feature engineering and model training, NEXUS takes a radically different route. Pre-trained on billions of real-world structured datasets and prediction tasks, it inherently understands how to extract critical signals from tabular formats.
Designed natively for structured data analysis, NEXUS introduces three key technical innovations: First, a deterministic architecture. Unlike probabilistic LLMs that can generate varying answers for identical queries, NEXUS yields consistent, highly reproducible results. Second, native tabular understanding. Having trained on billions of tables, it processes numbers, categories, dates, and unstructured text natively, eliminating manual feature engineering. Third, non-sequential reasoning. While typical AI models predict sequential data (such as the next word or pixel), NEXUS processes multi-dimensional relationships within enterprise tables simultaneously (e.g., correlating transaction frequencies, support tickets, and economic indicators to predict customer churn).
Existing approaches fall short because most enterprise data resides in tabular formats like ERPs, CRMs, and relational databases. Traditional ML workflows force data science teams into 3–6 month cycles of feature engineering, limiting the scale of predictions. Meanwhile, standard LLMs lose crucial numerical contexts during tokenization, leading to hallucinations and requiring massive guardrails to prevent incorrect outputs.
NEXUS overcomes these hurdles with dedicated architecture benefits, notably permutation invariance. It intuitively recognizes that changing column ordering does not alter the underlying semantic meaning of tabular data, distinguishing it fundamentally from standard text-based Transformers.
[AgentUpdate Depth Analysis] Historically, the AI Agent ecosystem has focused heavily on unstructured text processing via LLMs. However, the true bedrock of enterprise intelligence lies in structured tables and databases. The introduction of Large Tabular Models (LTMs) like NEXUS represents a paradigm shift. By bridging predictive AI with generative Agents, NEXUS empowers future enterprise Agents to execute highly complex, deterministic decision-making directly on transactional databases. Agents will no longer be limited to calling standard SQL queries or relying on brittle LLM-based reasoning on tables; they can leverage NEXUS as a 'quantitative brain' alongside the LLM's 'linguistic brain.' This hybrid intelligence architecture will dramatically lower the barrier for deploying autonomous, decision-making Agents in mission-critical environments like financial forecasting, supply chain optimization, and operational risk management.
