Hetzner Inference Experiment
Free hosted inference while the Inference API is in experimental status; Hetzner says it will notify users in advance by email if that status changes. Rate limits are applied per API key and the API returns HTTP 429 when any of them is exceeded, but as of 2026-09-23 the Inference API docs page no longer publishes the numeric request and token figures. Performance and availability are not guaranteed, requests can queue at peak, and the platform is not for production use.
The terms in full
Free hosted inference while the Inference API is in experimental status; Hetzner says it will notify users in advance by email if that status changes. Rate limits are applied per API key and the API returns HTTP 429 when any of them is exceeded, but as of 2026-09-23 the Inference API docs page no longer publishes the numeric request and token figures. Performance and availability are not guaranteed, requests can queue at peak, and the platform is not for production use.
- When it runs out Requests are rejected (429) until the window resets.
- Clock Explicitly an experiment with fewer guarantees than mature products and no committed end date; scope was already reduced once, so treat as volatile.
- Region Hetzner states that it hosts the Experiments platform on its own hardware in its own data centres. The Inference API pages do not state which countries or regions that hardware sits in, and the only EU-based description in Hetzner's material refers to the earlier OpenClaw Experiment rather than the public Inference API. No user-region restriction is stated.
Models
On the Inference API docs page (2026-09-17): Qwen/Qwen3.6-35B-A3B-FP8 and Qwen3.8-27B, both 262,144-token context with text and image input. The models endpoint is the definitive list and the selection changes during the experiment; GLM-5.2, DeepSeek-V4-Flash-0731 and Kimi-K2.7-Code were served at launch and dropped when Hetzner paused large models in August 2026.
Signup, payment, data
- Signup required; Hetzner account sign-in to the Hetzner Experiments Platform, then generate an auth token in the Inference tab at experiments.hetzner.com
- Card / payment not required: Hetzner's account-creation guide lists an email address, contact details and a billing address as the requirements, with adding a payment method only an optional later step
- Training Provider says it does not train on your prompts or outputs.
- Data terms The Inference API FAQ states: "We do only keep data that is necessary in order to track usage and (in the future) bill usage: request timestamps, token counts etc." It also states: "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)." Hetzner's blog post on the experiment adds, "We do not log any prompt data." No page states anything about using customer data for model training.
What the provider's page says
Every structured number above is backed by one of these quotes, checked verbatim against the page our crawler fetched. A field with no quote is shown as not stated.
- payment “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.” source
- limits “Timeframe Input Tokens Output Tokens 60s 4M 100k Additionally, we enforce request-level rate-limits as follows: Timeframe Requests 60s 10” source
- at exhaustion “If you exceed any of the rate limits, the API will respond with HTTP Status Code 429.” source
Evidence pages, re-read nightly
Change history
- Listing corrected: Hetzner Hetzner Inference Experiment our correction, not a provider change
- data treatment The Inference API FAQ states: "We do only keep data that is necessary in order to track usage and (in the future) bill usage: request timestamps, token counts etc." It also states: "We do not store the content of request and response and we do not plan on doing so in the future (unless some law requ
- Listing corrected: Hetzner Hetzner Inference Experiment our correction, not a provider change
- data treatment Hetzner's Inference API documentation states that the service keeps only what it needs for usage accounting: "We do only keep data that is necessary in order to track usage and (in the future) bill usage: request timestamps, token counts etc." The same page states, "We do not store the content of re
- Terms changed: Hetzner Hetzner Inference Experiment (free terms)
- free terms Free hosted inference while the Inference API is in experimental status; Hetzner says it will notify users in advance by email if that status changes. Rate limits are applied per API key and the API returns HTTP 429 when any of them is exceeded, but as of 2026-09-23 the Inference API docs page no lo
- Terms changed: Hetzner Hetzner Inference Experiment (free terms)
- free terms Free hosted inference while the Inference API is in experimental status; Hetzner says it will notify users in advance by email if that status changes. Rate limits are applied per API key and cover requests as well as input and output tokens, and the API responds with HTTP Status Code 429 when any li
- Listing corrected: Hetzner Hetzner Inference Experiment our correction, not a provider change
- data treatment Hetzner's Inference API FAQ says it keeps only the data needed to track usage and, in the future, bill usage, naming request timestamps and token counts. The same 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
- Terms changed: Hetzner Hetzner Inference Experiment (free terms, models detail, card)
- free terms Free hosted inference while the Inference API is in experimental status; Hetzner says it will email in advance if that changes. Published limits per API key: 10 requests per 60 seconds, 4M input tokens and 100K output tokens per 60 seconds, HTTP 429 past any of them. Performance and availability are
- models detail On the Inference API docs page (2026-09-17): Qwen/Qwen3.6-35B-A3B-FP8 and Qwen3.8-27B, both 262,144-token context with text and image input. The models endpoint is the definitive list and the selection changes during the experiment; GLM-5.2, DeepSeek-V4-Flash-0731 and Kimi-K2.7-Code were served at l
- card not required: Hetzner's account-creation guide lists an email address, contact details and a billing address as the requirements, with adding a payment method only an optional later step
- card required no
- Listing corrected: Hetzner Hetzner Inference Experiment our correction, not a provider change
- free terms Free hosted inference while the Inference API is in experimental status; Hetzner says it will email in advance if that changes. Published limits per API key: 10 requests per 60 seconds, 4M input tokens and 100K output tokens per 60 seconds, HTTP 429 past any of them. Performance and availability are
- models detail On the Inference API docs page (2026-09-17): Qwen/Qwen3.6-35B-A3B-FP8 and Qwen3.8-27B, both 262,144-token context with text and image input. The models endpoint is the definitive list and the selection changes during the experiment; GLM-5.2, DeepSeek-V4-Flash-0731 and Kimi-K2.7-Code were served at l
- card not required: Hetzner's account-creation guide lists an email address, contact details and a billing address as the requirements, with adding a payment method only an optional later step
- card required no
Evidence and scope
- Last verified
- 2026-10-09 (recheck due 2026-10-16)
- Evidence standard
- Provider-owned pricing, quota, limits, model or account pages only; third-party lists are never evidence.
- Data
- program.json, and the whole inventory at free-access.json (CC BY 4.0)
LLM Price Index lists what the provider publishes and links the page it came from. It does not test speed or quality, and a free route can change or end without notice; the change history above records what the nightly re-read confirmed.