predator-a3b-ngl6-ctx-1 — qwen3 on predator
45 calls across 3 cell(s); ~5.2 tok/s mean; p50 84.8s
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-ngl6-ctx64k | 5.4 | 5.3 | 5.9 | 1m13s | 15 |
| qwen3:30b-a3b-iq2m-think500-ngl6-ctx96k | 5.4 | 5.4 | 5.8 | 1m39s | 15 |
| qwen3:30b-a3b-iq2m-think500-ngl6-ctx131k | 4.8 | 4.8 | 4.9 | 1m38s | 15 |
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 | 5.2 | 1m21s | 9 | 5.5 | 1m10s | 1.05× |
| predator:llamacpp:qwen3:30b-a3b-i… | 3 | 5.3 | 1m50s | 9 | 5.5 | 1m28s | 1.04× |
| predator:llamacpp:qwen3:30b-a3b-i… | 3 | 4.7 | 2m0s | 9 | 4.8 | 1m36s | 1.03× |
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… | 5.5 tok/s 19.2s · n=4 | 5.5 tok/s 1m13s · n=4 | 5.2 tok/s 2m26s · n=4 |
| predator:llamacpp:qwen3:30b-a3b-i… | 5.6 tok/s 17.5s · n=4 | 5.4 tok/s 1m50s · n=4 | 5.3 tok/s 2m31s · n=4 |
| predator:llamacpp:qwen3:30b-a3b-i… | 4.7 tok/s 37.4s · n=4 | 4.8 tok/s 1m40s · n=4 | 4.8 tok/s 2m31s · 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… | 1031 | 891 | 1.16× |
| predator:llamacpp:qwen3:30b-a3b-i… | 1266 | 766 | 1.65× |
| predator:llamacpp:qwen3:30b-a3b-i… | 1250 | 715 | 1.75× |
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/79afe366.html Public benchmarks of the Weeyuga cluster. Run id: 79afe366-1055-4e45-adf6-593864a530e8. SHA ddbaaf46.