predator-a3b-ngl-matrix-1 — qwen3 on predator
42 calls across 3 cell(s); ~5.7 tok/s mean; p50 71.5s
Methodology
See A3B_AND_CPU_OVERNIGHT_2026-05-05
for the full procedure.
Reproducible at git SHA ddbaaf46.
Results
| Cell | tok/s mean | tok/s p50 | tok/s p95 | duration p50 | calls |
|---|---|---|---|---|---|
| qwen3:30b-a3b-iq2m-think500-ctx32k-ngl6 | 5.0 | 5.0 | 5.2 | 1m32s | 14 |
| qwen3:30b-a3b-iq2m-think500-ctx32k-ngl12 | 5.6 | 5.4 | 6.6 | 1m21s | 14 |
| qwen3:30b-a3b-iq2m-think500-ctx32k-ngl24 | 6.6 | 6.7 | 6.9 | 1m7s | 14 |
tokens per second — mean · p50 · p95
Cold start vs warm
Cold-start measurements are the first call into a model after it loads from disk; warm calls are everything after. The ratio shows how much of the deployment’s wall-time cost is one-time vs steady-state.
| Cell | cold n | cold tok/s | cold p50 | warm n | warm tok/s | warm p50 | warm/cold |
|---|---|---|---|---|---|---|---|
| predator:llamacpp:qwen3:30b-a3b-i… | 3 | 4.9 | 1m48s | 9 | 5.0 | 1m32s | 1.03× |
| predator:llamacpp:qwen3:30b-a3b-i… | 3 | 5.4 | 1m4s | 9 | 5.6 | 1m32s | 1.04× |
| predator:llamacpp:qwen3:30b-a3b-i… | 3 | 6.2 | 1m7s | 9 | 6.7 | 1m7s | 1.08× |
By prompt difficulty
Tokens per second by prompt class. hello is a trivial
one-line prompt; P-MEDIUM and P-HARD are the
deeper questions in the suite. The shape of the gap tells you whether
the model is bottlenecked on parsing or on generation.
| Cell | hello | P-MEDIUM | P-HARD |
|---|---|---|---|
| predator:llamacpp:qwen3:30b-a3b-i… | 4.8 tok/s 23.1s · n=4 | 4.9 tok/s 1m36s · n=4 | 5.1 tok/s 2m33s · n=4 |
| predator:llamacpp:qwen3:30b-a3b-i… | 6.2 tok/s 14.6s · n=4 | 5.3 tok/s 1m22s · n=4 | 5.3 tok/s 2m6s · n=4 |
| predator:llamacpp:qwen3:30b-a3b-i… | 6.6 tok/s 12.4s · n=4 | 6.7 tok/s 1m7s · n=4 | 6.5 tok/s 1m58s · n=4 |
Reasoning vs answer
Thinking models split their output into a hidden reasoning trace and a visible answer. The ratio shows how much of the budget the model spent thinking vs answering.
| Cell | reasoning chars | answer chars | reasoning / answer |
|---|---|---|---|
| predator:llamacpp:qwen3:30b-a3b-i… | 1264 | 657 | 1.92× |
| predator:llamacpp:qwen3:30b-a3b-i… | 1129 | 658 | 1.72× |
| predator:llamacpp:qwen3:30b-a3b-i… | 1095 | 736 | 1.49× |
Per-call timeline
Every call placed during this run, in order, colored by phase. Width is proportional to the call’s share of the cell’s wall-time. Hover any segment for the prompt id and tok/s.
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/b37836bf.html Public benchmarks of the Weeyuga cluster. Run id: b37836bf-d1a2-4d6b-a732-aff89da1fa07. SHA ddbaaf46.