ROBOTWAR.ioOSINT research running on the inference APIbeta test
Hetzner Inference API — Status
Independent measurement of https://inference.hetzner.com/api/v1 · one short question per model every 15 minutes · published 2026-10-23 16:19 CEST · all times on this page are Amsterdam (CET/CEST)
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.
Qwen/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
78.2 tokens/s
Also listed as Qwen/Qwen3.6-35B-A3B-FP8.
Active parameters, parameters read from the model name.
Qwen/Qwen3.6-35B-A3B-FP8
healthy
aliasQwen
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
0.33 s typical
throughput
53.2 tokens/s
Quantisation, active parameters, parameters read from the model name.
Same model, second name. This id is an alias of Qwen/Qwen3.6-35B-A3B. Both ids are still measured separately below: each is its own entry in the model list, and the timings of the two can differ.
Requests to this id are answered as Qwen/Qwen3.6-35B-A3B — in 392 of 392 successful measurements, going by the model name in each reply.
The endpoint's model list gives both ids the same creation timestamp (8 Oct 2026) and the same owner (hetzner).
First listed: this id on 16 Aug 2026, Qwen/Qwen3.6-35B-A3B on 9 Oct 2026.
Median reply: health check 0.33 s here, 0.25 s for Qwen/Qwen3.6-35B-A3B; working request 1.89 s here, 1.26 s for Qwen/Qwen3.6-35B-A3B.
Qwen/Qwen3.8-27B
healthy
Qwen
size
27B parameters
health check
0.47 s typical
throughput
30.1 tokens/s
Also listed as Qwen3.8-27B.
Parameters read from the model name.
Qwen3.8-27B
healthy
alias
size
27B parameters
health check
0.47 s typical
throughput
31.5 tokens/s
Parameters read from the model name.
Same model, second name. This id is an alias of Qwen/Qwen3.8-27B. Both ids are still measured separately below: each is its own entry in the model list, and the timings of the two can differ.
Requests to this id are answered as Qwen/Qwen3.8-27B — in 391 of 391 successful measurements, going by the model name in each reply.
The endpoint's model list gives both ids the same creation timestamp (6 Oct 2026) and the same owner (hetzner).
First listed: this id on 18 Aug 2026, Qwen/Qwen3.8-27B on 9 Oct 2026.
Median reply: health check 0.47 s here, 0.47 s for Qwen/Qwen3.8-27B; working request 3.66 s here, 3.82 s for Qwen/Qwen3.8-27B.
19 Oct 2026: Five fixed prompts sent to both ids at temperature 0 with a fixed seed came back word for word the same on four. On the fifth, Qwen/Qwen3.8-27B asked twice also differed from itself, and Qwen3.8-27B matched the repeat.
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
6641
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
1651
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.
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.
Click a model in the legend to hide or show it.
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.
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.
Click a model in the legend to hide or show it.
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
1113
11
0.25 s
0.36 s
52.2 s
23 ms
Qwen/Qwen3.6-35B-A3B
Working request
271
100
1.26 s
2.10 s
37.7 s
13 ms
Qwen/Qwen3.6-35B-A3B-FP8 (alias)
Health check
2320
11
0.33 s
0.71 s
55.7 s
30 ms
Qwen/Qwen3.6-35B-A3B-FP8 (alias)
Working request
564
100
1.89 s
4.43 s
64.0 s
19 ms
Qwen/Qwen3.8-27B
Health check
1075
11
0.47 s
2.33 s
58.5 s
42 ms
Qwen/Qwen3.8-27B
Working request
275
115
3.82 s
26.0 s
136.6 s
33 ms
Qwen3.8-27B (alias)
Health check
2133
11
0.47 s
1.85 s
59.7 s
43 ms
Qwen3.8-27B (alias)
Working request
541
114
3.66 s
25.8 s
177.9 s
32 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 hours100.0%1 week100.0%
Qwen/Qwen3.6-35B-A3B-FP8 (alias)
48 hours99.5%1 week99.9%
Qwen/Qwen3.8-27B
48 hours100.0%1 week98.6%
Qwen3.8-27B (alias)
48 hours99.5%1 week98.6%
Uptime over time
One point per day (Amsterdam) per model. A day nobody measured leaves a gap rather than a zero.
Click a model in the legend to hide or show it.
Availability, recent
Each cell is 30 minutes and shows the worst result in it, so a brief failure is never averaged away.
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 (2026-09-23 16:19 CEST onward)
Rate-limit headers
not returned by this endpoint — our own client-side budget is the limit