Concept card

AI Models Behind SAP Joule — Partner LLMs, SAP Models and Who Chooses

As of 2026-10-06

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is AI Models Behind SAP Joule?

Joule runs on a mix of third-party foundation models accessed through SAP's Generative AI Hub (SAP names OpenAI, Anthropic, Google Gemini, Meta Llama, Mistral) and SAP's own models (SAP-ABAP-1, SAP-RPT-1, SAP Domain Models); SAP does not publish a per-capability model map, and model choice by the customer happens only in custom-built AI.

The question

"Which model is Joule running on?" is asked at almost every SAP AI kick-off. The accurate answer in September 2026 has three parts: Joule uses several third-party foundation models plus SAP-built models; SAP names the vendors but does not publish which model serves which Joule capability; and the customer chooses models only when it builds its own AI on the SAP Business AI Platform.

Why it matters

Clients ask which model Joule uses; the defensible answer is 'several, chosen by SAP for standard capabilities' — and the useful follow-up is data-processing terms, hosting region and sub-processors.
Model choice is a real lever only in custom-built AI (Joule Studio, Generative AI Hub), where the customer selects, switches and pays for models at runtime.
Sovereignty-sensitive customers have named options — Mistral AI and Cohere on SAP's cloud infrastructure per Sapphire 2026 — that must be confirmed per region and capability.

Key points

Joule = multiple third-party foundation models + SAP models; SAP names vendors (OpenAI, PaLM 2/Gemini, Anthropic, Llama, Mistral) but publishes no per-capability model map.
Sapphire 2026: Claude 'among the foundation models' for Joule agents in HR, procurement and supply chain — not an exclusive or primary model.
Mistral AI and Cohere: sovereign model options on SAP's cloud infrastructure.
NVIDIA OpenShell secures the Joule Studio runtime; Google Cloud and Microsoft enable A2A interoperability; n8n in Joule Studio; Parloa in Service Cloud.
Palantir and Accenture are implementation partners for complex data migration — not Joule model providers.
Joule for Consultants uses Anthropic Claude on Amazon Bedrock via Generative AI Hub (AWS for SAP blog, Aug 2026).
SAP models: SAP-ABAP-1 (ABAP/CDS code), SAP-RPT-1/1.5/1.6 (tabular prediction), SAP Domain Models; Prior Labs TabPFN in Generative AI Hub.
Customers choose models only in custom-built AI (Joule Studio, Generative AI Hub); customer data is not used to train third-party models per SAP.

Terms used on this page

Generative AI Hub: Capability of SAP AI Core that gives governed access to third-party and SAP models, with orchestration services (templating, grounding, masking, evaluation); used for custom-built AI.
SAP-ABAP-1: SAP-built foundation model fine-tuned on the SAP S/4HANA ABAP and CDS code base, used for code explanation and generation.
SAP-RPT-1: SAP's relational/tabular foundation model that predicts on structured business data without model training or fine-tuning.
Sovereign model option: A model offered on SAP's cloud infrastructure to meet data-residency and jurisdiction requirements; Mistral AI and Cohere were named at Sapphire 2026.
SAP Domain Models: Models trained on SAP's own code and process knowledge, running 'under the hood' for S/4HANA and Ariba development rather than exposed directly as a chat model.
AWS Bedrock access path: One of the routes by which the generative AI hub reaches Anthropic Claude models — alongside direct SAP hosting — used for example by Joule for Consultants.
Underlying-model fact-sheet field: The field on an AI agent's SAP AI Agent Hub fact sheet naming its underlying model family — narrower than a public model map, but auditable per agent.
0.02 AI Units per action: SAP's flat 2026 price for every SAP-delivered autonomous-enterprise agent action, charged identically regardless of which model runs underneath.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →