Hetzner Inference API — Performance

Name: Hetzner Inference API — Performance
Creator: robotwar.io

Measured from robotwar.io Hetzner OSINT · generated 2026-10-17T01:47:22Z · window: last 30 days (all-time totals in summary)

← Back to the Hetzner OSINT feed · see also the API status report

Calls in window

3 106

6 887 all time

Success rate

97.3%

3 023 completed

Latency p50

24.4 s

p95 78.1 s

Output tokens/s

56.2

median of completed calls

Rate limited

0.0%

1 responses

Server errors

0.7%

22 incl. gateway 504

Timeouts

0.0%

0 calls

Parse errors

0.0%

0 calls

Tokens in

16 429 747

28 696 356 all time

Tokens out

4 752 302

8 775 801 all time

Attempts per success

1.0

1.0 = never retried

Retries

calls beyond the first attempt

Collection schedule

All times on this page are Amsterdam (CET/CEST). A missed slot means no collection started within the tolerance of a scheduled time.

Collection runs at 00:00, 08:00, 16:00 Europe/Amsterdam · next expected 2026-10-17 08:00 CEST · arrived 86 of 90 expected since 2026-09-17 · 3 did not finish

Last completed collection started 2026-10-16 16:00 CEST and took 21 min.

Missed: 2026-09-18 00:00, 2026-09-18 08:00, 2026-09-29 00:00, 2026-09-29 08:00

started	duration	status
2026-10-17 03:30 CEST	—	running (unfinished)
2026-10-17 00:00 CEST	23 min	failed
2026-10-16 16:00 CEST	21 min	ok
2026-10-16 08:00 CEST	29 min	ok
2026-10-16 03:30 CEST	47 min	ok
2026-10-16 00:00 CEST	27 min	ok
2026-10-15 16:00 CEST	26 min	ok
2026-10-15 08:00 CEST	21 min	ok
2026-10-15 03:30 CEST	21 min	ok
2026-10-15 00:00 CEST	25 min	ok
2026-10-14 16:00 CEST	25 min	ok
2026-10-14 08:00 CEST	19 min	ok

Model selection

Chosen from measurement: a model is used once the endpoint offers it and it has answered reliably, inside the response-time limit.

step	model	how	since	why
corroborate	Qwen/Qwen3.6-35B-A3B-FP8	held	2026-10-16	Qwen/Qwen3.6-35B-A3B leads by only 0.031 — under the 0.08 margin
deep	Qwen/Qwen3.6-35B-A3B-FP8	held	2026-10-16	still the best measured option
lens	Qwen/Qwen3.6-35B-A3B-FP8	held	2026-10-16	Qwen/Qwen3.6-35B-A3B leads by only 0.031 — under the 0.08 margin
synthesize	Qwen/Qwen3.6-35B-A3B-FP8	held	2026-10-16	still the best measured option
tag	Qwen/Qwen3.6-35B-A3B-FP8	held	2026-10-16	still the best measured option
trendwatch	Qwen/Qwen3.6-35B-A3B-FP8	held	2026-10-16	still the best measured option

Calls by status, per day

Stacked by outcome, Amsterdam days.

Speed over time

Latency percentiles cover completed calls; the dashed line is the 7-day mean.

Latency per model

Median latency per Amsterdam day.

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

Per model

Window: last 30 days. First/last seen span the whole log.

model	calls	success	p50	p95	max	tok/s	avg out	429	5xx	timeouts	first seen	last seen
Qwen/Qwen3.6-35B-A3B-FP8	3 106	97.3%	24.4 s	78.1 s	286.9 s	56.2	1 572	1	22	0	2026-08-16	2026-10-17

Model bench runs

Same prompt, same signals, every model.

when	model	parsed	latency	tokens in	tokens out
2026-08-16 19:26	Qwen/Qwen3.6-35B-A3B-FP8	yes	76.2 s	4 310	1 621
2026-08-16 19:25	Kimi-K2.7-Code	no	—	—	—
2026-08-16 19:17	GLM-5.2-NVFP4	yes	243.4 s	3 999	1 698
2026-08-16 18:57	DeepSeek-V4-Flash-0731	no	—	—	—

Per step

Repair calls are the single retry after an unparseable answer.

step	calls	success	p50	p95	avg in	avg out	repair calls
tag	1 168	98.9%	33.0 s	84.6 s	6 557	2 001	2
media	327	100.0%	689 ms	1413 ms	743	33	0
lens-cn	143	96.5%	20.9 s	60.9 s	4 844	1 318	17
lens-de	143	96.5%	19.0 s	61.1 s	4 787	1 309	18
events	137	96.4%	26.1 s	61.5 s	10 520	1 713	14
lens-pl	137	96.4%	21.4 s	62.1 s	4 896	1 359	12
lens-baltics	134	96.3%	25.6 s	93.3 s	5 140	1 644	1
trendwatch	132	90.9%	49.8 s	116.4 s	10 830	3 350	4
lens-ir	132	96.2%	24.5 s	56.8 s	4 772	1 463	6
synthesize	129	95.3%	30.8 s	104.1 s	5 667	2 059	0
lens-kp	129	96.1%	22.0 s	63.0 s	4 738	1 274	2
lens-wider	129	95.3%	22.4 s	74.4 s	5 124	1 513	1
lens-nl	128	95.3%	16.9 s	41.3 s	3 097	1 055	2
lens-by	128	96.1%	20.1 s	43.5 s	3 098	1 110	3
media-eval	10	100.0%	621 ms	1030 ms	746	32	0

Failures over time

Most recent failures

Last 20, newest first.

when	step	model	http	class	message
2026-10-06 22:22	trendwatch	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-06 06:06	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 22:21	trendwatch	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 22:14	synthesize	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 07:06	lens-nl	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:45	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:40	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:35	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:29	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:23	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:17	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:12	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-05 06:07	tag	Qwen/Qwen3.6-35B-A3B-FP8	504	gateway 504	504
2026-10-04 06:36	lens-wider	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.
2026-10-04 06:35	lens-wider	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.
2026-10-04 06:35	lens-wider	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.
2026-10-04 06:35	lens-wider	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.
2026-10-04 06:34	lens-wider	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.
2026-10-04 06:33	lens-by	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.
2026-10-04 06:32	lens-by	Qwen/Qwen3.6-35B-A3B-FP8	—	other	connection: Connection error.

Time of day

Amsterdam hours across the whole window.

Median latency — darker is slower

Failure share — darker is worse

3106 calls placed across the window

Model availability over time

One availability check per day from /models; calls per model per Amsterdam day; last 14 days. Watch this when a model disappears.

The endpoint currently lists 4 models: Qwen/Qwen3.6-35B-A3B, Qwen/Qwen3.6-35B-A3B-FP8, Qwen/Qwen3.8-27B, Qwen3.8-27B. Not offered on the latest check: DeepSeek-V4-Flash-0731 (last seen 2026-08-16); GLM-5.2-NVFP4 (last seen 2026-08-16); Kimi-K2.7-Code (last seen 2026-08-16). This timeline shows if and when they return.

	10-05	10-06	10-07	10-08	10-09	10-10	10-11	10-12	10-13	10-14	10-15	10-16	10-17
DeepSeek-V4-Flash-0731	×	×	×	×	×	×	×	×	×	×	×	×	×
GLM-5.2-NVFP4	×	×	×	×	×	×	×	×	×	×	×	×	×
Kimi-K2.7-Code	×	×	×	×	×	×	×	×	×	×	×	×	×
Qwen/Qwen3.6-35B-A3B-FP8	82	92	91	92	120	90	93	94	97	209	193	149	88
Qwen/Qwen3.6-35B-A3B	×	×	×	×
Qwen/Qwen3.8-27B	×	×	×	×
Qwen3.8-27B

offered and used successfully (number = ok calls) offered, every call failed offered, not used not offered on /models no availability check that day

Notable changes

2026-10-09 +Qwen/Qwen3.6-35B-A3B; +Qwen/Qwen3.8-27B

API surface over time

Capability flags are observed, never assumed.

taken	endpoint	sdk	models	change
2026-10-17 01:31	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-16 22:01	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-16 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-16 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-16 01:30	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-15 22:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-15 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-15 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-15 01:30	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-14 22:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-14 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-14 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-13 22:03	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-13 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-13 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-13 01:30	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-12 22:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-12 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-12 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-12 01:30	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-11 22:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-11 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-11 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-11 01:30	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-10 22:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-10 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-10 06:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-10 01:30	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-09 22:00	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-09 20:28	https://inference.hetzner.com/api/v1	2.49.0	4	no change
2026-10-09 14:00	https://inference.hetzner.com/api/v1	2.49.0	4	+Qwen/Qwen3.6-35B-A3B; +Qwen/Qwen3.8-27B
2026-10-09 06:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-09 01:30	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-08 22:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-08 14:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-08 06:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-08 01:30	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-07 22:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-07 14:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-07 06:02	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-07 01:31	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-06 22:01	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-06 14:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-06 06:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-06 01:30	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-05 22:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-05 14:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-05 06:01	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-05 01:30	https://inference.hetzner.com/api/v1	2.49.0	2	no change
2026-10-04 22:00	https://inference.hetzner.com/api/v1	2.49.0	2	no change

Observed on the latest snapshot

thinking can be switched off: yesgateway 504 observed: norate-limit headers exposed: no

Models offered

Qwen/Qwen3.6-35B-A3BQwen/Qwen3.6-35B-A3B-FP8Qwen/Qwen3.8-27BQwen3.8-27B

Per-model first and last call

model	first call	last call	calls (all time)
DeepSeek-V4-Flash-0731	2026-08-16	2026-08-16	6
GLM-5.2-NVFP4	2026-08-16	2026-08-16	6
Kimi-K2.7-Code	2026-08-16	2026-08-16	10
Qwen/Qwen3.6-35B-A3B-FP8	2026-08-16	2026-10-17	6 865

Rate-limit posture

Client budget: 10 requests / 60 s · observed 429 responses in window: 1 (0.0% of calls)

header	latest value
retry-after	1