pavilion-weeyuga-v3 — qwen2.5/qwen2.5-coder/qwen3/qwen3.5 on pavilion
TL;DR
5 of 5 Qwen3.5 35B IQ2_S task calls returned; 3/5 passed the format check. Median response time 27.35s. The route is not recorded, so this cannot be compared with the later local llama.cpp 30B run.
Archive run
This run is published for transparency. The site-grade is
archive-only — the run was meta-only,
didn’t complete cleanly, lacked documented methodology,
or had a methodology issue a newer run supersedes. Headline
numbers in this row should not be cited as current findings
without reading any caveat below and the underlying
run.md in the public archive.
Methodology
Task counts, output lengths and elapsed-time summaries are calculated from retained JSONL call records. No inference-throughput estimate is presented because the serving route is unverified.
See Public methodology and measurement limits
for the full procedure.
Recorded source revision (where available): 371ce70c.
Results
| Cell | task calls | successful | errors | format pass (task) | output tokens p50 | task duration p50 |
|---|---|---|---|---|---|---|
| qwen3.5:4b | 5 | 5 | 0 | 2/5 | 1124 | 28.77s |
| qwen3.5:35b-a3b-uncensored-iq1m | 5 | 5 | 0 | 3/5 | 1108 | 27.03s |
| qwen3.5:35b-a3b-iq2s | 5 | 5 | 0 | 3/5 | 1121 | 27.35s |
| qwen3.5:9b-q6k | 5 | 5 | 0 | 3/5 | 1124 | 27.13s |
| qwen3.5:9b-q4km | 5 | 5 | 0 | 3/5 | 1060 | 25.50s |
| qwen3.5:2b | 5 | 5 | 0 | 4/5 | 1060 | 25.77s |
| qwen3.5:0.8b | 5 | 5 | 0 | 4/5 | 1087 | 26.52s |
| qwen3.5:9b | 5 | 5 | 0 | 2/5 | 1104 | 26.86s |
| qwen2.5-coder:14b | 5 | 0 | 5 | — | — | — |
| qwen2.5-coder:3b | 5 | 5 | 0 | 3/5 | 305 | 24.79s |
| qwen3:14b | 5 | 2 | 3 | 2/2 | 845 | 290.17s |
| qwen3:8b | 5 | 4 | 1 | 3/4 | 854 | 218.77s |
| qwen3:4b | 5 | 5 | 0 | 0/5 | 2048 | 237.38s |
| qwen2.5:3b | 5 | 5 | 0 | 3/5 | 181 | 14.32s |
| qwen2.5-coder:1.5b | 5 | 5 | 0 | 3/5 | 271 | 10.60s |
| qwen2.5-coder:0.5b | 5 | 5 | 0 | 3/5 | 451 | 8.95s |
What these records can support
The April ledgers report similar ~40 tok/s rates for several Qwen3.5 sizes, but do not record the inference route. Those rates may reflect the serving path or a rate limit, so they are omitted here and should not be read as model throughput. This page shows retained task response counts, output lengths and elapsed times. The May Qwen3 30B result used a directly measured local llama.cpp server; its 5.88 tok/s mean covers 12 calls including cold starts, so it is not an equivalent comparison.
| Cell | cold n | cold tok/s | cold p50 | warm n | warm tok/s | warm p50 | warm/cold |
|---|---|---|---|---|---|---|---|
| pavilion:weeyuga:qwen3.5:4b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:35b-a3b-… | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:35b-a3b-… | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:9b-q6k | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:9b-q4km | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:2b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:0.8b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3.5:9b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen2.5-coder:14b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen2.5-coder:3b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3:14b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3:8b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen3:4b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen2.5:3b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen2.5-coder:1.5b | 2 | — | — | 4 | — | — | — |
| pavilion:weeyuga:qwen2.5-coder:0.5b | 2 | — | — | 4 | — | — | — |
Raw data
Every run gets its JSONL, log, summary, and metadata published. Clone the archive; re-run it; tell us where we got it wrong.
Cite
Margetic, S. et al. (2026). benchmarks.weeyuga.com/benchmarks/ad057f5b.html Public benchmarks of the Weeyuga cluster. Run id: ad057f5b-ed3f-4a95-a38e-361be310ffd6. SHA 371ce70c.