Hetzner is shutting down its free inference experiment on 8 December 2026

Since July 2026 you could call language models through an API at Hetzner free of charge, on servers in Germany and Finland. I introduced the offering in July and have measured it three times since. That chapter now has an end date. On 4 November 2026 Hetzner emailed everyone taking part in the experiment, from experiments@hetzner.com, subject “Hetzner Inference Experiment: Important changes and what’s next”.

The short version fits into one sentence. The free experiment ends on 8 December 2026, the offering becomes an official product, and today’s models will not be part of it.

What Hetzner told participants

The email lists five points. I am reproducing them here because they cannot be found anywhere else so far, and because they matter to anyone running the API.

No new access tokens. Hetzner is no longer issuing tokens for the experiment and explicitly advises against starting new implementations on it.

The Qwen models are going away. They are not part of the official product, because newer models have appeared since. Hetzner says it will reassess the Qwen series once new generations are released. Anyone using them should plan for an alternative.

Possible performance impact during the move. While the service migrates to new infrastructure, the experiment may not run as smoothly as usual.

A breaking change on 10 November 2026. The old base URL would only work until that date, after which every request returns an error. Hetzner notes that it normally avoids breaking changes, but that this one was needed for a major improvement to its inference infrastructure.

The end on 8 December 2026. After that the inference experiment is no longer available. Hetzner says it will announce the launch and the migration details for the official product through its official channels.

The email gives no launch date for the product. It only says development is nearly finished and that the launch is being prepared. So there may well be a gap between the end of the experiment and the start of the product. If you rely on the offering today, do not plan on something else being seamlessly available on 9 December.

Photo: panumas nikhomkhai / Pexels

The old URL no longer answers

I verified the announced breaking change on 14 November 2026. The old URL responds with HTTP 404 and a body that could hardly be shorter:

$ curl -s -o /dev/null -w "%{http_code}" \
    https://inference.hetzner.com/api/v1/models -H "Authorization: Bearer <token>"
404

$ curl -s https://inference.hetzner.com/api/v1/models -H "Authorization: Bearer <token>"
route not found

With the new URL the same request goes through normally:

$ curl -s -o /dev/null -w "%{http_code}" \
    https://inference.hetzner.com/v1/models -H "Authorization: Bearer <token>"
200

The only difference is api in the path. If you run an application that has been throwing errors since 10 November, that is your explanation. The error message is no help, because a 404 looks like a wrong model name or a typo. That is exactly why I am spelling it out here, so it can be found by someone searching for it.

In n8n it is enough to change the base URL in the OpenAI-type credentials, everything else stays the same. The other settings you need there are covered in a separate article. In the OpenAI SDK it is the base_url, and in hand-written calls it is the address in your code.

Almost none of this is public

This is the part that surprised me most while checking. The information sits in the email to participants. It is not in Hetzner’s own documentation. As of 14 November 2026:

• The Inference API documentation carries 24 July 2026 as its last change. It still names the old base URL https://inference.hetzner.com/api/v1, the one that stopped answering four days ago. No end date appears there.
• The Experiments platform overview, by contrast, was changed on 9 November 2026. Its list of completed experiments contains only OpenClaw. The Inference API is still described as a running offering, with no mention of the end.
• There is no product page for the future Inference API. The obvious addresses on hetzner.com return 404.
• On the Hetzner blog the most recent post on the topic is the review from 17 August 2026. Nothing has been added since.

For a business that means two things. Anyone who did not take part in the experiment had no way of knowing any of this. And anyone reading the official documentation today is given an address that no longer works. This is not a criticism of the decision to turn the experiment into a product, which I think is the right move. It is a reminder that with an offering carrying experimental status, you cannot count on the documentation keeping pace.

Photo: Rafael Minguet Delgado / Pexels

What still runs today, and how fast

Through the new URL the model list returned four entries on 14 November, covering two models that each carry a second name. Qwen/Qwen3.6-35B-A3B and Qwen/Qwen3.6-35B-A3B-FP8 are the same model, as are Qwen/Qwen3.8-27B and Qwen3.8-27B. The context window is unchanged at 262,144 tokens.

The speed is the more interesting part. I have measured this API since July with the same method, meaning five short runs for time to first token and a roughly 400-word explanation for throughput. Hetzner had warned about degradation during the move. I measured the opposite.

Measurement	24 July	23 August	21 September	14 November
Qwen3.6, time to first token	244 ms	257 ms	0.45 to 2.3 s	1.03 s
Qwen3.6, throughput	181 tok/s	75 tok/s	24 to 50 tok/s	105 tok/s
Qwen3.8, time to first token	not yet offered	364 ms	9.7 s to over 2 min	1.34 s
Qwen3.8, throughput	not yet offered	28 tok/s	4 to 7 tok/s	43 tok/s

Qwen3.8 throughput is six to ten times what it was in September. Back then the same model needed up to four and a half minutes for a 400-word explanation, now it takes about 17 seconds. Qwen3.6 writes at 105 tokens per second, roughly twice its September figure and 58 percent of its July value. Time to first token sits at about a second for both models, higher than in July but without the September outliers, where individual requests queued for more than two minutes.

Two caveats belong with this. These are snapshots from one afternoon, and in earlier runs the load varied noticeably within an hour. On top of that, far fewer people are likely to be hitting the service now that no new keys are issued. Even so, faster infrastructure matches what Hetzner gave as the reason for the URL change. The full comparison of the earlier runs is in my re-measurement after four weeks.

What to do before 8 December

If you have the API running somewhere, there are four steps.

First, switch the base URL if you have not already. Anything throwing errors since 10 November almost certainly comes down to this. Then search your projects for the old address, including environment variables, n8n credentials and scripts that only run occasionally. It is the monthly overnight job that otherwise surfaces the problem in January.

Second, find every place where the Qwen models are named in your code or configuration. Those names disappear when the product launches. If you keep the model name in a single place, the switch later is a one-line change.

Third, I would not build any new feature on this access path. For a prototype finishing within the next three weeks the offering is still good. For anything that needs to run in January it is the wrong foundation.

Fourth, you need a plan B for the time after 9 December. Because the interface follows the OpenAI format, switching is usually a matter of URL, key and model name. Which providers compute in Germany and the EU, what they cost and who offers a GDPR data processing agreement is covered in my comparison of EU-hosted providers. For tasks without personal data, another free tier will do in the meantime, and my piece on free AI APIs sets out what each provider may do with your prompts.

Photo: cottonbro studio / Pexels

What stays open

The questions that decide whether a business can use this are all still unanswered. What will the Inference API cost? Which models will it offer? Will there be a data processing agreement under Art. 28 GDPR, which was missing so far? Will there be an availability commitment? And when exactly does it launch?

The data processing agreement is the point that decides whether a playground becomes a tool for customer data. As long as it is missing, the API stays what it was during the experiment, namely good for prototypes, internal tools and anything without personal data. A price well below the American providers would still be a serious option for small and mid-sized companies, especially with servers in Germany and Finland. Whether it turns out that way will become clear with the launch announcement.

I will look at the product as soon as it is there and keep measuring with the same method. At that point this site will also say what it really costs and how fast it is.

If you are running an AI feature on this interface and want to work out what to move it to before December, get in touch.