Hetzner Quietly Launched a Free API for DeepSeek, GLM and Qwen Models

Hetzner launched a free Experiments Inference API in July 2026, offering developers an OpenAI-compatible endpoint for open-weight models like Qwen/Qwen3.6-35B-A3B-FP8, with no billing, no SLA, and no production guarantee. The service is hosted in Germany and Finland, positioning EU data residency as a key feature, but is not intended for production or sensitive workloads.

Hetzner has opened a free AI inference experiment for developers, and the important part isn't the price. It's whether a low-cost European host can become a serious LLM endpoint. Hetzner switched on its Experiments Inference API in July 2026, giving developers a way to run an OpenAI-compatible language model endpoint from Hetzner infrastructure without setting up their own GPU server. You create a token at experiments.hetzner.com, point an OpenAI-compatible client at Hetzner's base URL, and start sending chat completion requests. That's the hook. Not a dashboard. Not a grand launch. An endpoint. A Hetzner Community tutorial published on July 15 shows the working setup clearly: the base URL is https://inference.hetzner.com/api/v1, and the example model is Qwen/Qwen3.6-35B-A3B-FP8. The tutorial uses OpenCode, but the point is broader than one coding agent. Any tool that can speak to an OpenAI-compatible provider can be wired into Hetzner with a small configuration change. For developers, that is the whole attraction. You don't need to rewrite the application around a new SDK. You don't need to run vLLM yourself. You don't need to rent a GPU box and nurse it through driver updates, memory limits and model loading errors. You swap the endpoint and test. The experiment is real, but it is still an experiment Hetzner has not dressed this up as a finished cloud AI product, and you shouldn't treat it as one. The company's own Experiments material for OpenClaw, another service on the same platform, says the hosted LLMs are provided on a best-effort basis and can become temporarily unavailable during heavy demand. It also says experimental services are not meant for production environments. That is useful honesty. Independent writeups have found the same rough shape. Sliplane described Hetzner Inference in late July as an OpenAI-compatible API with no billing, no service-level agreement and no production guarantee. EMIT Solution, in a July 25 post, said the service was free during the test phase and noted the same practical limit: there was no data processing agreement at the time of its test, which matters if you want to send personal customer data through it. So you can use this for prototypes, internal tools, coding agents and experiments. You should be careful before pushing regulated or sensitive customer workloads into it. Free infrastructure is not the same as accountable infrastructure. Why Hetzner is trying this now Hetzner's move makes sense because inference has become one of the stickiest parts of the AI stack. Developers may compare model quality every week, but once an application is wired into an endpoint, that endpoint becomes part of the product's plumbing. Billing, latency, rate limits and data location all start to matter. Hetzner already has the customer base for this kind of test. The company is known among developers for cheap servers and blunt infrastructure, not for glossy enterprise software. That reputation helps here. If you already run VPS or bare-metal workloads on Hetzner, a hosted LLM endpoint from the same provider is easy to understand. There is also a European angle you can't ignore. Hetzner operates data centers in Germany and Finland, and many European teams have become more careful about where AI prompts and outputs travel. GDPR, customer contracts and internal security reviews all turn data location into a real procurement question. A German provider offering an OpenAI-compatible endpoint is not just copying the American AI cloud market. It is finding the part of the market where geography is a feature. Frankly, that is the strongest reason to pay attention. Hetzner doesn't need to beat OpenAI on model quality to make this interesting. It needs to make EU-hosted inference simple enough and cheap enough that developers keep it in their stack after the free test ends. The limits are just as clear. Hetzner is hosting open-weight models here, not building a frontier system of its own. The model list may change, the service may hit capacity problems, and the company has not published long-term pricing. Until that happens, nobody can compare it properly with OpenRouter, Together AI, Groq or the larger cloud providers. But the direction is plain. Hetzner is testing whether its old advantage in servers can carry over into AI inference. If enough developers show up, the free experiment becomes market evidence. If they don't, it stays what it is today: a useful, temporary endpoint with a very clear warning label. Also read: FlightAware Sues Kalshi Over Unauthorized Use of Its Flight Data https://startupfortune.com/flightaware-sues-kalshi-over-unauthorized-use-of-its-flight-data/ • Researchers used AI to build a Zoom hijacking exploit in one day https://startupfortune.com/researchers-used-ai-to-build-a-zoom-hijacking-exploit-in-one-day/ • Supermicro Shares Jump 8% as Margins Nearly Double Despite a Revenue Miss https://startupfortune.com/supermicro-shares-jump-8-as-margins-nearly-double-despite-a-revenue-miss/