Sovereign AI: AI infrastructure you control.
An open-weight language model on a dedicated GPU server in Germany, run by us. No model provider in between: it works offline and answers only our own systems. The rules are ours.
Who decides what your AI is allowed to know, say or do?
If your AI comes through another company’s API, that company decides: which models you get, what they may answer, what it costs and what happens to your data.
For many tasks that is fine. It stops being fine once your product, your research or your operations depend on it.
Sovereign AI means the operator decides. Which model runs, where it runs, how it behaves and what never leaves the machine.
What runs today
A private model behind a public character.
Our private model powers the conversational layer of SOLLE AI. The model has no public endpoint and no connection to any model provider. Our own systems reach it through a private tunnel.
The weights are fixed to one version and checked before they load, so the model only changes when we change it. How it behaves comes from our own configuration.
The setup in detail: running an LLM on a Hetzner GPU server with vLLM.
We now build the same for others: private AI on a dedicated GPU server in Germany, with your own model, your own rules and your data on your server.
- 01 Message Someone writes to SOLLE.
- 02 Gateway Arrives on a server we run.
- 03 Rules & limits Our rules and limits apply.
- 04 Private tunnel Restricted, no public access.
- 05 Open-weight model vLLM on our GPU server, offline.
- 06 Answer Back the same way.
Spec sheet
- Hardware
- Dedicated GPU server at Hetzner in Germany: NVIDIA RTX 4000 SFF Ada with 20 GB VRAM.
- Model
- Qwen3-14B, an open-weight model, in a compact 4-bit version (AWQ).
- Serving
- vLLM, running offline on the server. The model server keeps no request logs.
- Access
- Only our own systems, through a private tunnel. No public endpoint.
- Behaviour
- Set by our own instructions and limits.
- In use by
- Talk to SOLLE, the conversational layer of SOLLE AI.
- Running since
- July 2026.
What this is not
- A model we trained. It is an open-weight model that we configure.
- A public API. It answers only our own systems.
- AI without rules. It has rules, and they are ours.
Want this on a server of your own?
See Private AI