July 24, 2026
Server nerds enter the AI arena
Hetzner is working on LLM Inference
Cheap server giant tiptoes into AI and the comments are already fighting
TLDR: Hetzner, a popular low-cost hosting company, is testing a basic AI service with no guarantees, mostly to see if people even want it. Commenters split between excitement over a cheap Europe-based option and snarky nitpicks over mistakes, wording, and whether the whole thing is half-baked.
Hetzner — best known as the no-nonsense budget hosting company many internet builders swear by — has started experimenting with an AI chatbot-style service. Important word: experimenting. It’s free for now, has no promises, no reliability guarantees, and just one model to play with. In other words, this is less “grand launch” and more “we put it on the table, now let’s see who pokes it.” And the community absolutely did.
The loudest reaction was a mix of curiosity, cautious optimism, and instant nitpicking. One commenter jumped straight into grammar-police mode, calling out what they saw as a backwards explanation of the “enable_thinking” setting — a tiny detail, but exactly the kind of thing internet commenters love turning into a courtroom drama. Another person was harsher, basically saying the whole write-up was dragged down by “AI slop hallucinations,” which is the kind of insult that lands like a folding chair in a comment thread.
But not everyone came to boo. Some saw a bigger story: a European AI provider that could keep regulators happy and maybe give businesses a local alternative to the usual giant US platforms. Others were downright enthusiastic, cheering the idea of AI becoming so cheap it gets close to background utility pricing. There was even a casual rival recommendation tossed in — Infomaniak got a drive-by mention as the slightly more grown-up version of this idea. So yes, Hetzner’s AI test is small. But the reactions? Very large, very online, and already spicy.
Key Points
- •Hetzner has launched an experimental LLM inference service, but the article says it currently has no billing, no SLA, no production guarantee, and only one model.
- •The service is an OpenAI-compatible API running on Hetzner’s own infrastructure, allowing standard OpenAI client integration via Hetzner’s base URL.
- •The only model currently available is Qwen/Qwen3.6-35B-A3B-FP8, described as a 35B-parameter Mixture-of-Experts model with 3B active parameters, image support, a 262K context window, and FP8 weights.
- •The article includes informal test results reporting 153 ms median time to first token and 224 output tokens per second under limited conditions.
- •The article states that the model handled formatting, retrieval instructions, and an image correctly, but failed two simple arithmetic questions.