Hetzner is experimenting with LLM inference on its own infrastructure, offering an Open AI‑compatible API that currently serves only the Qwen/Qwen3.6‑35B‑A3B‑FP8 model, a 35‑billion‑parameter Mixture‑of‑Experts with a 262K context window. The service is free, has no SLA, and is meant to gauge demand and scaling.

Users create an API token in the Experiments dashboard, point any Open AI client at Hetzner’s base URL, and can toggle the enable_thinking option. In quick tests on July 23 2026 the API showed a median 153 ms time to first token and 224 output tokens per second on longer generations. The model handled text and image inputs well, though it stumbled on simple arithmetic.

Hetzner sees the experiment as a way to learn whether people want inference, how the system scales, and which features matter. The company’s strength lies in cheap GPU hardware and efficient operations; sharing GPU capacity via an API could turn idle hardware into revenue. However, its public bare‑metal lineup only offers NVIDIA RTX 4000 SFF Ada (20 GB) and NVIDIA RTX PRO 6000 Blackwell Max‑Q (96 GB) GPUs, which may limit it to smaller models unless it deploys larger B200/B300‑class clusters internally.

If the experiment leads to bigger GPU clusters and a broader model catalogue, Hetzner’s data‑center expertise, European positioning, and aggressive pricing could make it a serious inference competitor. For now the API is fast, free, and fun to try, with the next hardware announcement likely to reveal its true direction.