SAP debuts RPT-1, a foundation model (and open-source variant) for relational enterprise data

SAP has unveiled SAP RPT-1, its first large-scale foundation model for relational business data. The model is designed to handle structured, table-based datasets common in enterprise systems. It uses in-context learning, allowing users to provide a few example rows of data and receive predictions for new entries without the need for traditional model retraining.

The company describes RPT-1 as a “table-native” transformer that understands the schema and semantics of relational data. SAP said the approach reduces the complexity of building and maintaining multiple machine-learning models across business domains such as finance, supply chain and customer operations. By leveraging contextual examples rather than task-specific fine-tuning, RPT-1 aims to make predictive AI accessible to both business users and data scientists.

Notably, the RPT-1 family includes two commercial builds and an open-source variant named sap-rpt-1-oss, licensed under Apache-2.0 and available via GitHub and Hugging Face. The open variant is intended for research and community use, lowering the barrier for experimentation and innovation with enterprise-style tabular AI.

The announcement was made during SAP TechEd 2025 in Berlin. Analysts note that the open-source release expands the potential ecosystem beyond SAP’s own cloud services, making it relevant to developers, no-code tools and universities as well as enterprise buyers.

For enterprises running SAP’s ERP and analytics systems, RPT-1 could reduce the cost and time required to operationalise AI. The open-source variant could also enable smaller companies and academic teams to test and prototype tabular-AI use-cases without vendor lock-in.

While the concept aligns with industry efforts to extend foundation models beyond text and images into structured data, SAP emphasises that the model still depends on data quality, domain alignment and appropriate deployment practices. The open-source edition invites the community to validate, extend and integrate the model into broader toolchains.