Archive · vps-81 historical telemetry · local-mac/2026-04-12-qwen2_5_coder_3b-same-model-concurrency-sweep.html. Originally rendered 2026-04-12. Re-hosted from MyServers on 2026-05-06. Methodology and harness conventions may differ from what we use today; see /methodology.html for current standards. ← back to all benchmarks
Concurrency Sweep

Qwen2.5 Coder 3B Same-Model Sweep

This run selected the historically fastest local-Mac question types, repeated them into one focused packet, and then swept concurrency upward until throughput clearly stopped improving.

Peak combined system throughput: 16.057 tok/s at 4-up. Practical plateau start: 4-up.

Selection

Fast Question Types Chosen From History

QuestionCategoryAvg primary TPSAvg follow-up TPSAvg combined TPS
py_csv_parseparsing31.8926.9958.88
py_cli_argscli25.8226.6552.47
py_pydantic_modelvalidation25.7825.8551.62
py_typing_dataclasstyping26.2923.5549.84
py_file_scanfile_io23.8225.6749.49
Sweep

Concurrency Levels

ConcurrencyCombined system TPSPrimary system TPSParallel speedupTotal wall seconds
312.4076.8392.634293.458
416.0578.1732.516225.253
513.8627.6464.585248.877
615.7128.4253.369226.832
713.4887.5575.996270.096
Details

Per-Level Notes

3-up

Qwen2.5 Coder 3B same-model 3-up

Combined system throughput 12.407 tok/s, primary system throughput 6.839 tok/s, wall time 293.458s.

  • Parallel speedup: 2.634
  • Wall savings: 62.0%
  • Usable primary/follow-up: 20/20
4-up

Qwen2.5 Coder 3B same-model 4-up

Combined system throughput 16.057 tok/s, primary system throughput 8.173 tok/s, wall time 225.253s.

  • Parallel speedup: 2.516
  • Wall savings: 60.2%
  • Usable primary/follow-up: 20/20
5-up

Qwen2.5 Coder 3B same-model 5-up

Combined system throughput 13.862 tok/s, primary system throughput 7.646 tok/s, wall time 248.877s.

  • Parallel speedup: 4.585
  • Wall savings: 78.2%
  • Usable primary/follow-up: 20/20
6-up

Qwen2.5 Coder 3B same-model 6-up

Combined system throughput 15.712 tok/s, primary system throughput 8.425 tok/s, wall time 226.832s.

  • Parallel speedup: 3.369
  • Wall savings: 70.3%
  • Usable primary/follow-up: 20/20
7-up

Qwen2.5 Coder 3B same-model 7-up

Combined system throughput 13.488 tok/s, primary system throughput 7.557 tok/s, wall time 270.096s.

  • Parallel speedup: 5.996
  • Wall savings: 83.3%
  • Usable primary/follow-up: 20/20