[benchtest] Apple M1 Pro (14-core GPU) · 32GB · resident

[jonathanlourette]

benchtest output

### qwisp benchtest v0.3.9

| env | |
|---|---|
| chip | Apple M1 Pro (14-core GPU) |
| RAM | 32 GB |
| macOS | Version 26.6 (Build 25G72) |
| disk | APPLE SSD AP0512R, 494 GB, read 5.2 GB/s |
| power | battery, thermal nominal |
| model | Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16 |
| tier | resident C=256 → strict (lossless) |

| test | mode | TTFT | decode | tokens | stability |
|---|---|---|---|---|---|
| code-256 | strict | 1.7s | 53.6 tok/s | 256 | ok (1.00) |
| nl-256 | strict | 1.5s | 56.4 tok/s | 256 | ok (1.00) |
| long-600 | strict | 1.6s | 52.0 tok/s | 600 | ok (0.97) |

tail of long-600: `…l < 1), pivot `1` goes to index 0    - Result after partition: `[1, 3,`

_greedy; TTFT includes prefill; 25s total._

Notes (optional)

No response

[github-actions]

Thanks for the report! It parsed cleanly and is queued for bench/RESULTS.md. This issue will close automatically once the data PR is merged.