RPT-1: SAP Launches a Relational Foundation Model for the Enterprise (thenewstack.io)

SAP today unveiled RPT-1 (Relational Pre-Trained Transformer), its first “relational foundation model” designed natively for structured, tabular business data. The model will be offered in small and larger variants on SAP’s platform later this year and as an open-weight release on Hugging Face. SAP trained the open model on the Tremendous TabLib Trawl (T4) dataset — a 1.34 TB corpus of roughly 3.1 million tables — and enriched its commercial model with customer-provided tables. RPT-1 (formerly ConTextTab in SAP research) is optimized to perform regression, classification and forecasting over CSV-style tables via in-context learning and a simple API call, removing much of the traditional need for bespoke model training or heavy feature engineering. SAP is also launching an SAP-RPT Playground (CSV up to 2,073 rows × 50 columns) with examples for maintenance prediction, payment risk and churn. This matters because mainstream LLMs, trained on unstructured text, often struggle with numeric reasoning and tabular prediction; RPT-1 signals a shift toward domain-specific foundation models tuned for relational data. The open-weight release will accelerate community research in tabular foundation models, while the commercially enriched variant promises better real-world performance — albeit raising governance, privacy and reproducibility questions about customer-trained data. RPT-1 joins a small but growing field (startups like Kumo and academic labs) exploring how pre-trained relational models can simplify enterprise ML pipelines and speed deployment of analytics use cases.