Home / Hetzner

Hetzner Inference API (experiment)

docs.hetzner.com/general/company-and-policy/experiments/infeInference endpoint, experimental free APIinference/hetzner-inference-apiJSON compare report an error provider JSON

Free allowance: Free of charge while the Inference API remains in experimental status; two models, Qwen/Qwen3.6-35B-A3B-FP8 and Qwen3.8-27B, each with a 262,144-token context and text plus image input; limits per API key are 4M input tokens and 100k output tokens per 60 s and 10 requests per 60 s
Verdict: caveatusable, with a named limit or an unprobed read path
Cost: free tier
Identity: card?placeholder, not evidence; card required, phone unknown
Binding limit: request-rateper API key: 10 requests per 60 s, 4M input tokens per 60 s, 100k output tokens per 60 s; HTTP 429 beyond any of them
Evidence: reported6 sources; a third party said so
Freshness: stale on 2026-11-13checked 2026-10-29 (live); stale on 2026-11-13

Free tier	free of charge while the Inference API remains in experimental status; two models, Qwen/Qwen3.6-35B-A3B-FP8 and Qwen3.8-27B, each with a 262,144-token context and text plus image input; limits per API key are 4M input tokens and 100k output tokens per 60 s and 10 requests per 60 s unit tokens; metered on request-count; the vendor says it will email users before the experimental status changes; the model list changes during the experiment and GET /v1/models is definitive
Price	none stated
Limits	request-rate (binding) per API key: 10 requests per 60 s, 4M input tokens per 60 s, 100k output tokens per 60 s; HTTP 429 beyond any of them source
Binding	request-rate 10 requests per 60 s per key is the ceiling that decides throughput; the token windows are far above it
Identity	card (placeholder) unverified at the vendor: the docs page names no sign-up requirement. A third-party tracker reports that minting a token needs a Hetzner account whose payment verification can include a card charge or other routes
Egress	not published
Interface	rest endpoint: https://inference.hetzner.com/api/v1/chat/completions
Tooling	any OpenAI-compatible client at https://inference.hetzner.com/api/v1; endpoints /v1/models, /v1/completions and /v1/chat/completions
Composes with	Hetzner-hosted open-weight models behind a per-key ceiling of 10 requests per minute
Reliability	experimental service offered as is: performance and availability are not guaranteed, no backups, and the model list can change without a fixed schedule
Notes	the identity requirement is the open item: the kind is recorded as card because the only source says payment verification can include a card charge.
Evidence note	reported the rate limits and the free status are quoted from the vendor's docs page; the identity requirement is a third-party report because the key-minting page is a client-rendered application that could not be captured as text.
Parsed numbers	request-rate 10 requests

Evidence 6

date	source	what it proved
2026-10-28	docs.hetzner.com/general/company-and-policy/experiments/infe	the API is free while it stays experimental
2026-10-28	docs.hetzner.com/general/company-and-policy/experiments/infe	per API key, 4M input tokens and 100k output tokens per 60 seconds
2026-10-28	docs.hetzner.com/general/company-and-policy/experiments/infe	request-level limit of 10 requests per 60 seconds
2026-10-28	docs.hetzner.com/general/company-and-policy/experiments/infe	Qwen3.8-27B is served with a 262,144-token context and image input
2026-10-28	docs.hetzner.com/general/company-and-policy/experiments/infe	no availability guarantee for the experimental service
2026-10-28	mvalentsev.github.io/awesome-free-ai-coding/providers/hetzne	third-party report on what getting a key requires

Unknown values 3

$.egress.metered egress metering was not published
$.per_object_cap.stated the field does not apply to this kind of offer
$.retention.stated the field does not apply to this kind of offer