Hetzner's Experiments platform: a free, honest AI inference experiment that could actually scare the big clouds

Hetzner launched its Experiments platform (experiments.hetzner.com) on July 24, 2026: free, hosted AI experiments running on Hetzner's own hardware in its own EU data centers (Nuremberg, Falkenstein, Helsinki). Two things are live: an Inference API and a hosted OpenClaw assistant. No billing, no SLA, no production guarantee, no backups. Deliberately. Here is what is actually good about it, and what should keep you skeptical.

The Inference API is an OpenAI-compatible REST endpoint (https://inference.hetzner.com/api/v1). You grab an API token from the Experiments dashboard, point any OpenAI SDK at it, done. Right now there is exactly ONE model: Qwen/Qwen3.6-35B-A3B-FP8. That is a 35-billion-parameter Mixture-of-Experts model with only 3 billion active parameters per token. It takes text and images, has a 262,144-token context window, uses FP8-quantized weights, and is Apache 2.0 licensed from Alibaba's Qwen team. The FP8 weights are roughly 38 GB, so real VRAM usage sits above that depending on context and cache.

Early hands-on tests (Sliplane, July 23) measured about 153 ms median time-to-first-token and roughly 224 output tokens per second. Fast for what it is. But the reviewer also noted the model failed two very simple arithmetic questions - a small, slightly flawed LLM. Hetzner's own docs admit quality "is probably not as good as the major providers." Fine for experiments, not for production.

Now the critical part, because there is plenty to be critical about.

First, the limits are real. One model, period. No xAI, no Anthropic, no OpenAI models, no embeddings, no image generation. Per-key rate limits: 3M input / 60k output tokens per 60 seconds, 500M input / 5M output per 24h, then HTTP 429. No backups exist, so your configs can vanish. High-demand periods may make the whole thing unavailable. Hetzner explicitly says do not run production on this.

Second, the hardware question is unanswered. Hetzner's public GPU lineup tops out at the RTX 4000 SFF Ada (20 GB) and RTX PRO 6000 Blackwell Max-Q (96 GB) - workstation GPUs, not the B200/B300-class multi-GPU monsters needed for the really big open models (think GLM-5 at 754B parameters across eight GPUs). Hetzner may run different internal hardware behind the API, but nobody knows. If this stays a one-or-two-small-model experiment, it stays a curiosity.

Third, the OpenClaw experiment has sharp edges. Your instance is publicly reachable, protected only by a UI password. Hetzner states plainly it has no influence over the OpenClaw code and accepts no liability for its behavior. No file-system access, no out-of-band shell, no backup functionality - deleting an instance permanently discards its storage. That is fine for playing, uncomfortable for anything you care about.

Here is why I still lean positive on this. Hetzner is doing something rare: shipping an honest experiment with the limitations printed on the box. No hype, no fake "AI experience" marketing - just a free, OpenAI-compatible API on sovereign European infrastructure with open-weight models under permissive licenses. The rate limits are generous for a free tier. The 262K context is genuinely large. And the strategic signal matters: if anyone can turn open-weight inference into a brutal low-margin commodity product, it is the company that already runs some of Europe's cheapest, most efficiently operated data centers. This is Hetzner testing whether spare GPU capacity can become a real revenue line.

The verdict: as a product, it is not there yet - one smallish MoE model, no guarantees, no backups, workstation-class GPU ceiling. As a direction, it is one of the most interesting things in European AI this year. Try it, break it, give feedback - but do not put your startup on it yet.