Hetzner GPU: Free Testing Isn’t So Simple

What Hetzner Actually Offers for Models

As of August 2026, the main catch is that Hetzner describes GPUs as dedicated servers rather than the usual cloud GPU instances with quick start and stop. This is clear in the official Hetzner GPU matrix and the GEX44 press release: the primary public product is not a cloud VM but a dedicated server.

The GEX44 comes with NVIDIA RTX 4000 SFF Ada, 20 GB VRAM, 64 GB RAM, and an Intel i5-13500. In Hetzner's materials, it’s directly positioned as a machine for open-source models, including Mixtral scenarios via API. Community instructions mention Gemma 3 12B, Phi 4 14B, and Mistral Small 24B as a reasonable workload class for the GEX44.

The high-end tier looks entirely different: the GEX131 with NVIDIA RTX PRO 6000 Blackwell Max-Q, 96 GB VRAM, 256 GB RAM, and an Intel Xeon Gold 5412U. In the same materials, it’s positioned for heavier models like Llama 3.3 70B. Regarding locations, the official matrix lists GEX44 in Falkenstein and GEX131 in Nuremberg and Falkenstein.

On the pricing front, a key detail is that Hetzner's press release for the GEX44 states €184.00 per month plus a one-time setup fee of €79.00, including IPv4. The Hacker News discussion also mentions €1.5980 per hour for the GEX44, but that's a community discussion figure, not the main rate card.

Where the Idea of a Free Experiment Breaks Down

A free test here may not be free at the point of registration. In the original discussion, one participant wrote that after signing up, they were asked to top up their balance with €25 to access the free experiment.

I’d treat this as an important signal but not as a confirmed rule for GPUs. None of the provided official Hetzner documents directly link GPU servers to a free trial with a minimum €25 top-up. So the surprise is possible, but the exact mechanism isn’t proven by these sources.

Engineering-wise, this changes expectations: you need to plan not as a Lambda-style one-off evening run, but as renting specific hardware with access conditions, a setup fee, and potential account verification. For experimenting with Ollama and Libre WebUI, it could still be interesting, especially if you need a private sandbox for open-source LLMs. But the word "free" is dangerous here: it too easily masks the deposit, fixed rental, and time required for manual setup.

My Take on This Story

The core idea is sound: the GEX44 covers a small class of local LLMs, while the GEX131 provides headroom for the 70B segment. These are not toy specifications, especially when 20 GB and 96 GB VRAM set the boundaries of what can fit without offloading struggles.

But the marketing phrasing around free experiments without precise terms easily trips up developers. The most annoying bug here isn’t in CUDA or the model but in expectations: you think you’re doing a quick test of an open-source model, but you end up in a process of purchasing access to a dedicated server.

We previously covered free access to Pony Alpha via OpenRouter, allowing safe model testing without upfront payment. Hetzner also launches free experiments with open models, but with a €25 deposit — a similar approach with an additional barrier.