Hetzner is serving free LLM inference from an unannounced experiments page

Sources: write-up, HN 49033087
Summary: A write-up published 2026-07-24 reports that Hetzner is running an experimental inference service on its experiments platform, exposing an OpenAI-compatible API that serves Qwen 3.6 35B at no charge with no SLA and no production guarantee. The author includes dashboard screenshots, a working code sample, and a measured 153 ms median time to first token from testing on 2026-07-23, and states plainly that they have no insider information and that nobody at Hetzner described a plan. Hetzner has published no announcement, and this run could not resolve a public Hetzner page describing the service.
Why it matters: A European host running its own datacenters offering token-billed inference would change the choice EU teams currently make between US APIs and self-hosting, which is why an unannounced experiment is worth tracking rather than reporting as a launch.
Follow-up: Watch for a Hetzner announcement, pricing, and whether the endpoint survives the experiment.