Skip to main content

Hetzner Inference Provider Guide

Hetzner Inference is a Tier-2 catalog provider: its integration is one JSON file (src/lib/providers/catalog/hetzner-inference.json) rather than hand-written code. That file is the source of truth for this page.

Verification status: this entry is docs-verified only, not yet live-verified. Hetzner's GET /models needs a key, so the model ids come from its public docs page and the roster has not been checked. No account was created and no API key was used to build the entry. evidence.liveMatrix is null (see Verification status below).

Key Facts​

  • Provider id: hetzner-inference (alias hetzner)
  • Protocol: OpenAI-compatible (/chat/completions) — the docs page says "The Inference API provides access to open-source large language models via an OpenAI-compatible REST API." (docs)
  • Base URL: https://inference.hetzner.com/api/v1
  • Default model: Qwen/Qwen3.6-35B-A3B-FP8
  • Models in catalog: 2 — the two rows of the docs page's Available Models table, with ids copied from the table cells
  • Vision: true on both catalog rows — the table's Modalities cell reads "Text, Image" for each
  • Streaming: not declared (false)
  • Tool calling: not declared (false)
  • Tools while streaming: not declared (false)
  • Structured output: not declared (false)
  • Structured output + tools together: not declared (false) — no combined probe was possible without credentials
  • Embeddings: not declared (false)
  • Thinking: not declared (false)
  • Service status: the docs page's note reads "This service is provided for experimental purposes only and is offered "as is"." and "Performance and availability are not guaranteed, especially during periods of high demand." The same note reads "No backups are created, so you are responsible for backing up your own configurations and should not use the platform for production environments."
  • Billing: free-tier — the docs FAQ says "As long as the Inference API remains in experimental status, it is free of charge."
  • Key format: none declared

Quick Start​

1. Get an API token​

  1. Visit: https://experiments.hetzner.com/inference — the page is client-rendered; its public JavaScript bundle labels the token area "Tokens" and the button "Create API Token" (sign-in was not attempted while building this entry)
  2. Create an API token and copy it. The docs' curl examples (https://docs.hetzner.com/general/company-and-policy/experiments/inference/) use the header Authorization: Bearer <YOUR_TOKEN>
  3. Billing as the vendor states it: "As long as the Inference API remains in experimental status, it is free of charge." (FAQ, same docs page)
  4. Set HETZNER_INFERENCE_API_KEY in your .env file

2. Configure​

export HETZNER_INFERENCE_API_KEY=your-api-token
export HETZNER_INFERENCE_MODEL=Qwen/Qwen3.6-35B-A3B-FP8   # optional — overrides the default model
export HETZNER_INFERENCE_BASE_URL=https://inference.hetzner.com/api/v1   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "hetzner-inference",
  model: "Qwen/Qwen3.6-35B-A3B-FP8",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider hetzner-inference

Per-request credentials:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "hetzner-inference",
  credentials: {
    hetznerInference: { apiKey: process.env.HETZNER_INFERENCE_API_KEY },
  },
});

Models​

ModelContextVisionNotes
Qwen/Qwen3.6-35B-A3B-FP8 ⭐262,144yesNeuroLink default. Type "Causal LM + Vision (MoE, 35B total / 3B active)"; named in the docs page's four code examples, including its image example.
Qwen3.8-27B262,144yesFallback. Type "Dense"; Modalities "Text, Image".

The figures come from the Available Models table of https://docs.hetzner.com/general/company-and-policy/experiments/inference/ (retrieved 2026-09-29): Context Length "262,144 tokens" and Modalities "Text, Image" on both rows. The same page says "Only a selected list of models is available. This selection will change during the experimental phase." and "The response of the models endpoint is definitive."

models.defaultContextWindow (32,768) and models.defaultMaxOutputTokens (4,096) are placeholders the vendor does not publish. They apply to model ids outside the catalog, and the output figure applies to both catalog rows, which carry no maxOutputTokens.

Fallback order when the default is unavailable: Qwen3.8-27B.

Licenses, as the page's Open-Source Model Licenses table lists them: rows "Qwen3.6-35B-A3B" and "Qwen3.8-27B", each with Developer "Alibaba Cloud (Qwen Team)" and License "Apache 2.0". The page also says "By using this API, you acknowledge the applicable license terms of the underlying model(s)."

Rate limits​

The docs FAQ says "Current limits are applied per API key:" followed by a table with Timeframe "60s", Input Tokens "4M" and Output Tokens "100k". A second table, introduced by "Additionally, we enforce request-level rate-limits as follows:", lists Timeframe "60s" and Requests "10". The FAQ continues: "If you exceed any of the rate limits, the API will respond with HTTP Status Code 429."

errorRules maps HTTP 429 to NeuroLink's rate-limit error class and HTTP 401 to its authentication class.

Data handling​

The docs FAQ says "We do not store the content of request and response and we do not plan on doing so in the future (unless some law requires us to do so)."

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Hetzner Inference — docs-verified only, not live-verified:

ProbeResult
Rosterunauthenticated GET https://inference.hetzner.com/api/v1/models, HTTP 401, 2026-09-29 — no key used; the model ids are taken from the docs page's Available Models table, and the roster has not been checked
Auth-failure shapethe same unauthenticated GET /models returned HTTP 401, content-type application/json; charset=utf-8, body {"error":"unauthorized"}; an unauthenticated GET /chat/completions on the same day answered HTTP 404, content-type text/plain; charset=utf-8, body "404 page not found". No POST request was sent
Billingpublic docs FAQ, 2026-09-29: "As long as the Inference API remains in experimental status, it is free of charge. Should this status change, we will notify you in advance via email with detailed information."
Capabilitiestext is true because the docs page shows chat-completion examples in Python and curl; the other capability flags are false (not declared), and no request exercising them was sent
Live capability sweepnot run. evidence.liveMatrix is null.

Troubleshooting​

SymptomWhat the vendor states or what was observed
HTTP 401Observed 2026-09-29: an unauthenticated GET /models answered HTTP 401 with {"error":"unauthorized"}. NeuroLink reads the token from HETZNER_INFERENCE_API_KEY
HTTP 429Docs FAQ: "If you exceed any of the rate limits, the API will respond with HTTP Status Code 429." Limits are listed under Rate limits
Model ids differ from the table aboveDocs page: "This selection will change during the experimental phase." and "The response of the models endpoint is definitive."

See also​