Hetzner Inference API — Status

Independent measurement of https://inference.hetzner.com/api/v1 · one short question per model every 15 minutes · published 2026-10-17 03:47 CEST · all times on this page are Amsterdam (CET/CEST)

← Back to the Hetzner OSINT feed · see also the performance report

Models listed

4 seen in window

Answering now

responded on the last cycle

Best uptime

99.8%

of models with enough measurements

Fastest health check

0.25 s

median reply, short request

Measurements

7082

2844 cycles

The models

Size and precision are read from each model's own name; anything else needs a cited source. Weights are parameters times bytes per parameter — serving needs more than that.

Qwen3.6-35B-A3B

healthy

Qwen

size: 35B total, 3B active
routing: mixture of experts · 9% active per token
health check: 0.25 s typical
throughput: 54.8 tokens/s

Active parameters, parameters read from the model name.

Qwen3.6-35B-A3B-FP8

healthy

Qwen

size: 35B total, 3B active
routing: mixture of experts · 9% active per token
served at: FP8 (1 byte per parameter)
weights: ≈ 35 GB
health check: 0.39 s typical
throughput: 63.5 tokens/s

Quantisation, active parameters, parameters read from the model name.

Qwen3.8-27B

healthy

Qwen

size: 27B parameters
health check: 0.50 s typical
throughput: 27.2 tokens/s

Parameters read from the model name.

Qwen3.8-27B

healthy

size: 27B parameters
health check: 0.53 s typical
throughput: 43.4 tokens/s

Parameters read from the model name.

What is measured

Synthetic probes against the live endpoint — not this section's research workload, which is reported separately in the performance report. Two request sizes are sent, and they are never mixed in a chart or a percentile.

request	kind	measurements	what is sent
Health check	ping	5372	A fixed one-line question with a per-request code in it, capped at 64 tokens. Sent to every model every 15 minutes. Identical every time, so this line is a true like-for-like comparison over time.
Working request	work	1349	A short passage to summarise in about 60 words plus three key terms, capped at 400 tokens. Sent hourly. It produces roughly ten times the output of the health check, which is why it is slower and why it is charted separately.

Response time — health check

The identical short request, every measurement, at the time it was taken. Timeouts are excluded: a timeout measures the deadline, not the model.

Qwen/Qwen3.6-35B-A3B (589 measurements)Qwen/Qwen3.6-35B-A3B-FP8 (2252 measurements)Qwen/Qwen3.8-27B (560 measurements)Qwen3.8-27B (1971 measurements)

Each point is one measured reply, plotted at the time it was taken. A break in a line is a stretch nobody measured, not a fast reply.

Response time — working request

The larger request, on the same scale idea but its own chart. It asks for about ten times the output, so its times are not comparable with the health check above.

Qwen/Qwen3.6-35B-A3B (145 measurements)Qwen/Qwen3.6-35B-A3B-FP8 (550 measurements)Qwen/Qwen3.8-27B (145 measurements)Qwen3.8-27B (511 measurements)

Each point is one measured reply, plotted at the time it was taken. A break in a line is a stretch nobody measured, not a fast reply.

Speed by request size

The last column normalises the two: milliseconds per output token. When those agree, the difference in raw seconds is the size of the request, not the endpoint.

model	request	n	tokens out (median)	p50	p90	slowest	per output token
Qwen/Qwen3.6-35B-A3B	Health check	589	11	0.25 s	0.38 s	52.2 s	23 ms
Qwen/Qwen3.6-35B-A3B	Working request	144	100	1.33 s	2.37 s	37.7 s	13 ms
Qwen/Qwen3.6-35B-A3B-FP8	Health check	2252	11	0.39 s	0.74 s	55.7 s	35 ms
Qwen/Qwen3.6-35B-A3B-FP8	Working request	549	100	2.15 s	4.49 s	64.0 s	22 ms
Qwen/Qwen3.8-27B	Health check	560	11	0.50 s	2.98 s	58.5 s	45 ms
Qwen/Qwen3.8-27B	Working request	145	115	4.13 s	32.5 s	136.6 s	36 ms
Qwen3.8-27B	Health check	1971	11	0.53 s	2.60 s	59.7 s	48 ms
Qwen3.8-27B	Working request	511	114	4.19 s	66.3 s	179.8 s	37 ms

Uptime

Two windows on purpose. 48 hours answers 'is it healthy now'; a week answers 'can I plan on it'. A single average hides a model that broke yesterday.

Qwen/Qwen3.6-35B-A3B

48 hours99.5%1 week99.7%

Qwen/Qwen3.6-35B-A3B-FP8

48 hours99.5%1 week99.9%

Qwen/Qwen3.8-27B

48 hours94.0%1 week95.6%

Qwen3.8-27B

48 hours99.0%1 week97.9%

Uptime over time

One point per day (Amsterdam) per model. A day nobody measured leaves a gap rather than a zero.

Qwen/Qwen3.6-35B-A3B (9 measured days)Qwen/Qwen3.6-35B-A3B-FP8 (30 measured days)Qwen/Qwen3.8-27B (9 measured days)Qwen3.8-27B (30 measured days)

Availability, recent

Each cell is 30 minutes and shows the worst result in it, so a brief failure is never averaged away.

Qwen/Qwen3.6-35B-A3B

Qwen/Qwen3.6-35B-A3B-FP8

Qwen/Qwen3.8-27B

Qwen3.8-27B

2026-10-15T02:00Z2026-10-17T01:30Z

healthyslowwrong answerrate limitedfailedunservablenot offerednot measured

Availability, every model

Window: 30 days. Speed is reported separately, per request size — a percentile over both sizes mixed together describes neither.

model	listed	uptime	answered	tokens/s	measurements
Qwen/Qwen3.6-35B-A3B	yes	99.7%	99.9%	54.8	735
Qwen/Qwen3.6-35B-A3B-FP8	yes	99.8%	99.9%	63.5	2806
Qwen/Qwen3.8-27B	yes	95.9%	95.9%	27.2	735
Qwen3.8-27B	yes	88.5%	88.5%	43.4	2806

What went wrong

The breakdown behind each uptime figure.

model	failures	breakdown
Qwen/Qwen3.6-35B-A3B	2	wrong answer 1, failed 1
Qwen/Qwen3.6-35B-A3B-FP8	5	failed 4, wrong answer 1
Qwen/Qwen3.8-27B	30	failed 30
Qwen3.8-27B	324	failed 322, rate limited 1, unservable 1

Incidents

Two or more consecutive bad cycles. A single blip is not an incident.

model	from	length	mostly	measurements
Qwen/Qwen3.8-27B	2026-10-17T00:16:03Z	15 min	failed	2
Qwen/Qwen3.8-27B	2026-10-14T10:16:03Z	17 min	failed	2
Qwen3.8-27B	2026-10-14T10:16:03Z	20 min	failed	2
Qwen3.8-27B	2026-10-13T22:01:52Z	15 min	failed	3
Qwen3.8-27B	2026-10-13T17:32:04Z	29 min	failed	3
Qwen/Qwen3.8-27B	2026-10-13T17:31:04Z	15 min	failed	2
Qwen/Qwen3.8-27B	2026-10-13T12:16:02Z	30 min	failed	3
Qwen3.8-27B	2026-10-08T16:01:04Z	44 min	failed	4
Qwen3.8-27B	2026-10-07T19:48:06Z	42 min	failed	4
Qwen3.8-27B	2026-10-07T18:33:05Z	42 min	failed	4
Qwen3.8-27B	2026-10-07T15:16:03Z	122 min	failed	9
Qwen3.8-27B	2026-10-07T14:16:03Z	32 min	failed	3
Qwen3.8-27B	2026-10-07T13:01:06Z	44 min	failed	4
Qwen3.8-27B	2026-10-07T11:16:03Z	44 min	failed	4
Qwen3.8-27B	2026-10-07T10:16:03Z	29 min	failed	3
Qwen3.8-27B	2026-10-07T06:03:03Z	43 min	failed	4
Qwen3.8-27B	2026-10-07T05:01:06Z	44 min	failed	4
Qwen3.8-27B	2026-10-07T02:01:02Z	30 min	failed	3
Qwen3.8-27B	2026-10-07T01:16:06Z	15 min	failed	3
Qwen3.8-27B	2026-10-07T00:31:03Z	14 min	failed	2

Changes to the model list

A model withdrawn from the list is not an outage — watch this before reading one.

2026-10-09 added Qwen/Qwen3.6-35B-A3B, added Qwen/Qwen3.8-27B (4 listed)

Were we looking?

Measurement coverage per hour.

48 of the last 48 hours were measured. Amber is a partly-measured hour and pale grey means nobody was looking — neither means anything was down.

About this endpoint


Service status	Beta. Hetzner offers this endpoint under Experiments, so it is a trial service rather than a general-availability product — expect the model list, and the numbers below, to move.
Endpoint	https://inference.hetzner.com/api/v1
Probe interval	every 15 minutes, per model
Gateway cut-off	300.0 s — a request past this is ended by the gateway
Streaming	not used, so latency below is the whole response, not time to first token
Window	30 days (2026-09-17 03:47 CEST onward)
Rate-limit headers	not returned by this endpoint — our own client-side budget is the limit