Hetzner's free inference experiment caps output at 60,000 tokens a minute

Point the OpenAI client at https://inference.hetzner.com/api/v1, pass a Hetzner token, and every other line of your code stays as it was ^[4][5]. There is no GPU to rent, no weights to download and no inference server to configure; Hetzner handles serving, batching and scaling ^[16]. Because the compatibility runs both ways, switching away means changing that one base URL again ^[5]. The token is displayed once at creation ^[6], so put it in the secret store immediately.

The model explains why this is cheap for Hetzner to run at all. Qwen3.6-35B-A3B-FP8 is a mixture-of-experts model with 35 billion parameters and 3 billion active per token ^[8], so about 8.6 percent of the weights are engaged for any given token ^[1]. The dev.to write-up describes the FP8 quantisation as halving memory against FP16 without major quality loss ^[9].

The published limits decide whether it does more than fill a notebook. Sixty thousand output tokens every 60 seconds ^[10] is 1,000 output tokens a second ^[2]. Sliplane.io measured 224 output tokens a second on a single stream ^[15], so roughly four simultaneous generations at that rate consume the whole allowance ^[3]. Input is roomier: 3 million tokens a minute ^[10] is about eleven requests at the full 262,144-token context ^[4], and 50 permitted input tokens for every output token ^[5]. Summarising long documents fits that shape.

Sliplane's other figure, a 153 ms median time to first token across seven short requests, came with the tester's caution that it was a single-client snapshot and not an SLA ^[15]. The test is dated July 23, 2026, the day before Hetzner announced the endpoint ^[6]. To see those numbers on your own workload you would need short prompts, one client at a time, a network path close to Germany or Finland, and roughly the traffic the endpoint carried before anyone knew it existed. Hetzner's docs say performance and availability are not guaranteed, especially during high demand ^[13].

On the one-model catalogue, Hetzner's documentation said: "We are starting with a single high-quality model to validate the platform, gather feedback, and ensure reliability. Additional models will be added based on demand and operational experience" ^[14]. Publishing the ceiling and the caveat before anyone builds on it is good practice. The rate limits, the as-is language and the instruction to keep production environments off the platform are all in Hetzner's documentation ^[10][13].

The service is free while it keeps experimental status, and Hetzner says that if the status changes it will notify users in advance by email with detailed information about billing changes ^[11]. It has not published future token prices ^[12]. So the evaluation costs nothing and the production case cannot be costed yet. No backups are created, and the docs tell users not to use the platform for production environments ^[13].