Fundamental's NEXUS tabular foundation model launches on Amazon SageMaker JumpStart

Fundamental's NEXUS tabular foundation model became available on Amazon SageMaker JumpStart on 2026-06-03, giving AWS customers a one-click deployment path to a pre-trained model designed specifically for structured-data prediction tasks rather than text generation. The release puts a tabular foundation model on the same procurement footing as the LLMs that have dominated SageMaker's catalog for the last two years.

What's new

• NEXUS is described by AWS as "a foundation model developed by Fundamental and built for tabular data prediction" — distinct from generative LLMs in both training objective and target workload.
• The model is pre-trained on "billions of real-world prediction tasks across structured datasets" and "natively processes numbers, categories, dates, and unstructured text without manual feature engineering."
• NEXUS is designed to produce "consistent, reproducible results for each individual prediction" — a deterministic property that matters in regulated workflows like credit scoring or fraud detection where audit trails track model decisions.
• Distribution is via SageMaker JumpStart, AWS's marketplace for foundation models that customers can deploy directly into their own AWS accounts, sit behind VPC endpoints, and call from SageMaker pipelines without leaving their security boundary.

Context

Tabular foundation models are a relatively young category. Most existing tabular ML still uses XGBoost or LightGBM with hand-crafted feature engineering for each dataset — a workflow that hasn't been displaced by general-purpose LLMs because language models are inefficient at the kinds of dense numeric reasoning that structured-data work requires. A handful of research groups have argued since 2024 that a separate "tabular foundation" approach — large pre-training on heterogeneous structured datasets, then few-shot or zero-shot inference on a new table — could collapse much of that feature-engineering burden. NEXUS is one of the more visible commercial productizations of that idea.

AWS surfacing NEXUS on JumpStart also fits a broader pattern: financial services and operations teams want to use the same procurement, security review, and deployment plumbing for tabular models that they already use for LLMs. NVIDIA's own developer blog this week argued the same thesis from the other direction, naming Revolut, Mastercard, Adyen, and Stripe as institutions building transaction foundation models on structured data. Cloud marketplaces are the channel that turns those bespoke deployments into off-the-shelf products.

Why it matters

The SageMaker JumpStart distribution is the practical lever here. Bank, insurance, and large-enterprise ML teams have historically struggled to get novel models through internal procurement without months of vendor review; JumpStart short-circuits that by hosting the model under the customer's existing AWS contract and security posture. That changes the customer profile a tabular foundation model can reach — from research-forward fintechs to mainstream regulated buyers — and it does so without Fundamental having to negotiate enterprise deals one at a time.

The more important signal is that AWS is willing to feature a tabular-only foundation model alongside its LLM catalog. JumpStart's curation acts as a soft endorsement of the category. If NEXUS sees meaningful usage on AWS, expect other tabular-foundation startups to follow it onto JumpStart, Azure AI Foundry, and Vertex Model Garden — and expect the long-running XGBoost-vs-deep-learning debate inside enterprise ML teams to tilt further toward pre-trained tabular models for greenfield work.