Private AI
Private LLMs.
Uncensored.
Yours.
Your own AI on a dedicated GPU server in Germany. Self-hosted open-weight LLMs (often called open-source AI), installed, adapted and run by us. No model provider in between, no filter you didn’t choose. Prompts and answers are processed on your server.
Uncensored doesn’t mean unaccountable.
Hosted AI answers under someone else’s policy. A private model answers under yours.
We run open-weight models, including variants freed of their refusals by abliteration: for research, security work, sensitive analysis, creative work, or because an adult should decide what their own tools may discuss.
You set the rules and carry the responsibility. Illegal use stays illegal, and we don’t build for it.
Why Europe
Sovereignty needs a place: a server in the EU.
Your instance runs in Germany, at Hetzner, a German provider with its own data centres in Nuremberg and Falkenstein. No big cloud platform in between, no model provider on another continent.
My goal is a European home for independent AI, with uncensored models in the hands of people who use them responsibly.
Who it is for
-
Companies
Private models and agents on your own server, connected to your own data. No API whose terms can change overnight.
-
Governments & public sector
A dedicated machine whose model makes no outbound calls. Open weights in a fixed version, so you can check exactly what runs.
-
Private individuals
Your own AI, for people who don’t want their questions to become someone else’s data.
What you get
Your server. Your model. Your rules.
- Hardware
- A dedicated GPU server at Hetzner in Germany: 96 GB of GPU memory (VRAM, GEX131) for large models and fine-tuning, or 20 GB (GEX44) for fast mid-size models. Rented in your name or ours.
- Models
- Open-weight models of your choice, uncensored variants included. Swap them when better ones appear.
- Interface
- A private chat interface and an API for your own tools. On request, we build the agents and automations that put your AI to work.
- Data
- Prompts and answers are not logged unless you decide otherwise. We sign a data processing agreement (DPA) with every client. More: self-hosted AI and the GDPR.
- Access
- Yours. We keep one admin login for maintenance until you want full handover. Then root access and every key are yours alone.
- Operation
- We look after the machine (updates, load, disk, health), not your conversations.
- Cost
- Priced per project, plus the server as a separate monthly rent at Hetzner. Both are clear before anything starts.
Model catalogue
Open-weight and uncensored LLMs, by GPU size.
Open-weight models we deploy, from a fast model for a 20 GB GPU to a 284-billion-parameter flagship. Uncensored variants have their refusals largely removed. Every licence allows private commercial use and fine-tuning.
- Qwen3-14B (AWQ): 20 GB, 24 GB, 96 GB
- Gemma 4 26B-A4B: 20 GB, 24 GB, 96 GB
- SuperGemma 4 26B: 20 GB, 24 GB
- Qwen3.6 35B-A3B: 24 GB, 96 GB
- Qwen3.8 27B: 24 GB, 96 GB
- Gemma 4 31B: 96 GB
- gpt-oss-120b: 96 GB
- DeepSeek V4 Flash: 96 GB
SuperGemma 4 26B, uncensored
Google Gemma 4A community variant tuned to feel less censored in everyday use.
Qwen3.6 35B-A3B, uncensored
QwenMany users at once, agents and tools, long context. Fast, because only 3 billion parameters are active at a time.
Qwen3.8 27B, uncensored
QwenReasoning, coding and agents, with image input. The strongest dense model of its size.
Gemma 4 31B, uncensored
Google Gemma 4Research, sensitive analysis, writing and long documents. One of the most willing open models under 40 billion parameters.
DeepSeek V4 Flash, abliterated
DeepSeekMaximum knowledge and quality for single users or small teams.
Willingness to answer: the independent UGI leaderboard (10 = answers everything), otherwise the author’s own test. Before going live, every model is tested against your own questions. More on GPU sizes: LLM VRAM requirements and the VRAM calculator.
Adapt the model
Fine-tuned on your data. Abliterated if you want.
An off-the-shelf model knows the internet. Yours should know your work. We adapt it on the 96 GB server, with your data.
- 01
Train
We fine-tune the model (LoRA/QLoRA) on your documents, language and field, so it answers like someone who knows your work. Your data never leaves the server.
- 02
Remove refusals
We take the refusal behaviour out of the model’s weights (abliteration) and can shift its tone and priorities.
- 03
Test
Before anything goes live, we test the model against your own questions, side by side with the original.
How to start
A small lab. A few projects at a time.
I’m Sven Lehnert and have been building software since 2008. Uncensored AI is powerful, so I take on a few projects at a time and build each one properly.
- 01
Write to me
A few sentences on what you want to build and where it has to run.
- 02
A short first call
Free. We see whether we are a good fit.
- 03
Private AI Architecture Briefing
Optional. 45 minutes with me on your case: a clear recommendation for model, hardware and architecture, plus a written summary. €490, credited if we build together.
- 04
Build
Server set up at Hetzner, model installed and adapted, tested against your questions, handed over.
Questions
Private LLM hosting, briefly answered.
What is uncensored AI?
An open-weight large language model (LLM) whose refusals have been largely removed by abliteration or fine-tuning. It answers what it is asked. You set the rules and carry the responsibility.
Is this a local or on-premise LLM?
In terms of control, yes. The model runs on a dedicated GPU server in Germany instead of in your building, with no cloud API in between. Rented in your name if you like, with root access and every key yours on request: on-premise control without buying hardware.
Is a self-hosted LLM GDPR-compliant?
Self-hosting takes the model provider out of the chain. Prompts and answers are processed on your server in Germany and are not logged unless you decide otherwise. We sign a data processing agreement (DPA) with every client.
Is it like a private ChatGPT?
In daily use, yes: a private chat and an API for your own tools, with the model you choose, uncensored variants included. The difference: open weights on your own server instead of a provider’s cloud.
What does it cost?
The first call is free. The optional Private AI Architecture Briefing costs €490 and is credited if we build together. Projects are priced individually; the server is a separate monthly rent at Hetzner.
Own the intelligence you depend on.
Get in touch