# Soofi S: Germany’s sovereign open-source foundation model (introduction)

> Soofi S is a 30B Mixture-of-Experts model trained on 27T tokens by a German research consortium under the IPCEI-CIS programme, plus the open-source soofi-trainer agent stack — what it means for sovereign AI delivery.

For AI agents: curated site map at [https://www.devolute.org/llms.txt](https://www.devolute.org/llms.txt). Canonical HTML (preferred for search engines): [https://www.devolute.org/en/insights/open-source/soofi-sovereign-german-ai-model](https://www.devolute.org/en/insights/open-source/soofi-sovereign-german-ai-model).

- Language: en
- Published: 2026-07-15
- Markdown URL: https://www.devolute.org/en/insights/open-source/soofi-sovereign-german-ai-model.md

## A German-made open model enters the sovereignty conversation

Soofi S ships model weights and an open agent stack — evaluation and integration still take engineering.

Part of **[open-source knowledge introductions](/en/insights/open-source)**. Strategy hub: **[digital sovereignty agency](/en/agency/digital-sovereignty-agency)**. Related: **[open-source self-hosted AI platform](/en/agency/open-source-self-hosted-ai-platform)**.

## What Soofi S is

**[Soofi](https://www.soofi.info/)** ("Sovereign Open Source Foundation Models") is a German research project launched in September 2025 and funded through the **IPCEI-CIS** programme under the Federal Ministry for Economic Affairs and Energy (funding ID 13IPC040A). Its first release, **Soofi S**, is a **30B-parameter Mixture-of-Experts model** trained on roughly **27 trillion tokens**, with a focus on German and English text. The consortium brings together German research institutes, universities, and start-ups, with backing referenced by the AI Bundesverband, Deutsche Telekom, and the Federal Ministry ([source: NXT EU on X](https://x.com/nxt4eu/status/2077343621011935648)).

The stated focus is **industrial AI**: technical and regulatory documents, code generation, agentic systems, and enterprise-specific applications — deployable on European cloud and edge infrastructure rather than only via third-party inference APIs. As of publication, Soofi S is in **practical testing with industry partners**; general release for direct use has not yet happened, and interested organisations are directed to contact the consortium directly.

## soofi-trainer: the open agent stack around the model

Alongside the model track, the consortium publishes **[soofi-project](https://github.com/soofi-project)** on GitHub, including **[soofi-trainer](https://github.com/soofi-project/soofi-trainer)** (MIT-licensed) — an agentic system that guides users through **LLM specialization**: from use-case analysis to method recommendation (RAG, fine-tuning, LoRA/QLoRA, SFT, DPO) and on to a managed training job. Technically it composes patterns our clients already run in production:

- **LangGraph ReAct orchestration** — an interaction agent, an advisor agent, a dataset agent, and a training agent, coordinated via A2A and exposed through AG-UI/SSE streaming.
- **MCP-first integration** — a Vector MCP server for knowledge-base search (`search_documents`, `list_metadata`) and a Training Gateway exposed as MCP for job management, plus an MCP Inspector for debugging.
- **RAG on Weaviate** — markdown knowledge documents with YAML metadata, ingested automatically with SHA-256 change detection, embeddings configurable per provider (OpenAI, Ollama `bge-m3`, or others).
- **Pluggable inference backends** — OpenAI by default, with drop-in profiles for Ollama, LM Studio, NVIDIA Triton, and vLLM via LiteLLM for on-prem H200 deployments.
- **Voice I/O** — Whisper-based STT and OpenAI/Piper TTS, German and English, with push-to-talk in the UI.
- **Operational scaffolding** — Docker Compose with domain-scoped sub-files, Portainer, Grafana/Prometheus, and MinIO object storage.

There is also an **AAS stack** (BaSyx Asset Administration Shell registries and browsers) and an **EDC stack** (Eclipse Dataspace Connector provider/consumer, with an MCP-wrapped consumer) — both pointing at Industry 4.0 and data-space interoperability use cases, not just chat.

## Why this matters beyond the model card

A capable open-weight model is a headline; what determines whether it is **usable in production** is the same as for any self-hosted stack: model serving choices, **[Kubernetes platform engineering](/en/agency/kubernetes-platform-engineering)**, **[Terraform / OpenTofu infrastructure](/en/agency/terraform-opentofu-infrastructure)**, and a **[RAG / retrieval](/en/agency/open-source-rag-retrieval)** layer that is actually evaluated against your documents — not the vendor's demo corpus. Soofi's own reference stack already assumes this: MCP servers, LangGraph agents, and a training gateway that treats fine-tuning as a managed job rather than a one-off script.

For organisations evaluating Soofi S specifically for **[digital sovereignty](/en/agency/digital-sovereignty-agency)** reasons — EU-hosted inference, auditable weights, no forced dependency on a single US cloud vendor — the model closes one gap, but governance, data pipelines, and operational ownership remain engineering work.

## Related introductions

**[Digital sovereignty — governance and data control](/en/insights/open-source/digital-sovereignty-governance-and-data-control)** · **[European digital sovereignty & EuroStack](/en/insights/open-source/european-digital-sovereignty-eurostack)** · **[Introduction to LLM agents](/en/insights/open-source/introduction-llm-agents)** · **[Introduction to RAG & retrieval](/en/insights/open-source/introduction-rag-retrieval)**