pi-hetzner-inference-provider

A Pi extension that registers the Hetzner Experiments Inference API as a first-class provider: OpenAI-compatible open-weight models (Qwen, DeepSeek, GLM, Kimi) served on Hetzner's own infrastructure. The model catalog mirrors the hetzner provider in models.dev — the same catalog OpenCode ships — so models, context/output limits, modalities and thinking levels behave exactly like they do in OpenCode.

Install

# via npm
pi install npm:pi-hetzner-inference-provider

# via git
pi install git:github.com/ipmanlk/pi-hetzner-inference-provider

# local development
pi install ./pi-hetzner-inference-provider

Setup

1. Create an API key:

   • Log in at https://experiments.hetzner.com with your Hetzner account.
   • On the left-hand sidebar, under APPS, select Inference.
   • Click the Create API Token button in the top-right corner.

2. Set the token as an environment variable:

   echo 'export HETZNER_API_KEY="<your-token>"' >> ~/.bashrc
   source ~/.bashrc

   HETZNER_VLLM_API_KEY and HETZNER_INFERENCE_API_KEY are accepted as aliases. You can also run /login hetzner inside pi and paste the token there.

3. Start pi and pick a model:

   pi
   # /model hetzner/Qwen/Qwen3.6-35B-A3B-FP8

Available models

| Model id | Context | Max output | Modalities | Thinking | | ------------------------- | --------- | ---------- | ----------------- | --------------------- | | DeepSeek-V4-Flash-0731 | 512 000 | 384 000 | text | off/low/high/max | | GLM-5.2-NVFP4 | 512 000 | 131 072 | text | off/high/max | | Kimi-K2.7-Code | 262 144 | 262 144 | text, image | always on | | Qwen/Qwen3.6-35B-A3B-FP8| 262 144 | 262 144 | text, image | always on |

Context and max-output figures come from models.dev limit for the hetzner provider; modalities from its modalities. The catalog is a static snapshot — there is no live refresh, so new models appear when this package is updated.

Thinking levels

Thinking control matches models.dev's reasoning_options for the provider:

• DeepSeek-V4-Flash-0731 — reasoning_effort in none/low/high/max, selectable in pi via /thinking (off → none, low → low, high → high, max → max).
• GLM-5.2-NVFP4 — reasoning_effort in none/high/max (off → none, high → high, max → max). GLM's reasoning at high/max can be very verbose on this experimental deployment — pick off for fast responses.
• Kimi-K2.7-Code and Qwen/Qwen3.6-35B-A3B-FP8 — no reasoning_options in models.dev, so no thinking parameter is sent at all: the models reason natively and their thinking is displayed, exactly like in OpenCode.

How it works

• Provider id hetzner, base URL https://inference.hetzner.com/api/v1.
• API openai-completions (OpenAI Chat Completions) with vLLM compatibility flags: max_tokens (not max_completion_tokens), no store flag, system role only.
• Thinking is sent as OpenAI-style reasoning_effort (DeepSeek, GLM), mirroring the models.dev catalog; Kimi/Qwen get no thinking parameter.
• Catalog is a static snapshot of the models.dev hetzner provider (no live refresh; update the package for new models).
• Cost is tracked at $0 — the Inference API is free while in experimental status.
• A User-Agent header (pi-hetzner-inference-provider/<version>) is sent on model requests.

Troubleshooting

• Stuck on “Working…”: the Inference API is experimental and can be very slow — first requests to a model can take a minute or more while it warms up, and requests can hang when the service is under load. Retry, or switch model. If the request eventually fails, run /login hetzner to re-check the token.
• GLM-5.2-NVFP4 too slow: GLM generates very verbose reasoning before answering, which on this experimental deployment can take minutes. Set the thinking level to off (/thinking off, or model id suffix :off in /model) — the reasoning_effort: none request disables reasoning completely.

License

MIT