OCTOO AI

OCTOO AI is the independent AI lab of Sven Lehnert in Las Palmas de Gran Canaria.

Our offer is Private AI: an open-weight LLM, standard or uncensored, that we install, fine-tune and operate for you on a dedicated Hetzner GPU server in Germany. That means the 20 GB GEX44 class while Hetzner still rents it, or the 96 GB GEX131 in Falkenstein. Hetzner offers the 24 GB GEX45 only in Helsinki, so we set that one up on request. You get a private chat interface and an API. Request logging is off by default, and we sign a GDPR processing agreement (Art. 28) with every client.

We run Qwen3-14B AWQ ourselves, with vLLM 0.30.0 on a 20 GB RTX 4000 SFF Ada. Measured: 28.8 tokens per second for one stream, about 54 ms to the first token, and 56.4 tokens per second combined with two requests at once.

A model without someone else's refusals still needs rules of its own. Before a model we adapt goes live, it answers the client's own questions side by side with the original. Illegal use stays illegal, and we do not build for it.

What you find here

• Open-weight LLMs we offer, standard and uncensored: the models in our catalogue, with file sizes, willingness scores and licences.
• Abliteration: papers, leaderboard and models: research on removing refusals from a model's weights, and on what else the edit changes.

Read more

• Which open and uncensored LLMs fit 20, 24 or 96 GB of VRAM
• Running an LLM on a Hetzner GPU server with vLLM: our setup, measured
• How a model loses its "no": abliteration, Heretic and the UGI score
• Self-hosted LLM and GDPR: what your own server solves, and what it doesn't

Private AI puts models like these on a server of your own: get in touch.