Self-hosted LLMs such as DeepSeek for companies in Germany

What a self-hosted LLM is

A self-hosted LLM is a language model that runs on hardware your company controls, in your server room or at a hosting provider under your contract. Open-weight models make this possible: the maker publishes the trained weights, and serving software such as vLLM or Ollama answers your applications through an API. DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, under the MIT license, with instructions for vLLM and SGLang on its model page.

Open weights are narrower than open source. You receive the trained model, usually without the data it was trained on.

Why companies in Germany ask about DeepSeek

The DeepSeek app and a DeepSeek model on your own server are two different things. On June 27, 2025, the Berlin data protection commissioner reported the app to Apple and Google as illegal content, because it transfers personal data of its users to China (press release). That concerns the app and DeepSeek’s own cloud. Weights that run on your server have no connection to DeepSeek.

Few companies in Germany use open models so far. In a Bitkom survey of 603 companies with 20 or more employees, published on September 9, 2026, 76 percent of the companies that use AI used ChatGPT, 8 percent Llama, 2 percent DeepSeek, 1 percent Qwen and 0 percent Mistral (Bitkom). Bitkom adds that many tasks do not need the largest and newest model and run on smaller systems at lower cost.

Licenses of open-weight models, checked on September 24, 2026

“Open” means something different for each maker. The table lists the latest release I checked per model family and what a company in the EU has to note.

Model family	Release checked	License	Note for a company in the EU
DeepSeek (China)	DeepSeek-V4.1-Flash, September 10, 2026	MIT	Commercial use allowed. A self-hosted copy sends no data to DeepSeek.
Qwen (Alibaba, China)	Qwen3.6-27B, April 2026	Apache 2.0	Commercial use allowed. Some larger Qwen models are available only as a cloud service.
Mistral (France)	Mistral Large 3, December 2, 2025	Apache 2.0	Commercial use allowed, from a maker in the EU.
gpt-oss (OpenAI, USA)	gpt-oss-120b and gpt-oss-20b, August 2025	Apache 2.0	Commercial use allowed. The 120b model fits on a single 80 GB GPU.
Llama 4 (Meta, USA)	Llama 4 Scout and Maverick, April 2025	Llama 4 Community License	A company with its principal place of business in the EU receives no rights to the multimodal models.

Sources: DeepSeek, Qwen, Mistral, gpt-oss, Llama 4 use policy, license audit of August 2026. A maker can change the license with its next release, so the assessment checks it again.

What I set up around the model

The model is one part of the work. The other part decides who may ask it what, and where its answers go:

• the server: your own rack, a hosting provider in Germany or a cloud region in Frankfurt, with a second site where the model must not go down,
• serving software with a token per application and a log of the requests,
• a connection to your systems through an API or an MCP server with a short list of tools, the way AI agents reach the API in my client work,
• the agents that use the model, under the rules on AI agents under the same rules as humans.

The agents can run on an open-source control plane, as described on Open-source AI agents. The rules around coding agents are on Agent harness engineering.

Who I am

I am Michael Wutzke, based in Frankfurt, and have worked in IT and media for more than twenty years. At the Frankfurt-based company Blocksize Capital I was Head of Decentralized Finance and Node Operations, then CIO. I ran blockchain nodes in several data centers there and moved the infrastructure from Azure to another cloud with failover. Today I help a client rebuild its platform with Claude and Codex under written rules, and I teach at Claude Hacker House. Beyond the models, I am interested in open-source agent software that runs on a company’s own server. Details: Career stages.