#3611 [Benchmark] Mac Studio M5 Ultra 96GB / Qwen3.8-Flash-Next IQ3_XXS / spec2 vs spec3 / strata-qwen38-thinking-v1-draft.1

本文

Results-only benchmark PR per BENCHMARK_PROTOCOL.md §Submit and review.

**What:** two standard-benchmark runs (`strata bench run`, protocol `strata-qwen38-thinking-v1-draft.1`) on one Mac Studio (Apple M5 Ultra, 96 GiB shared memory), engine 0.1.47-metal source build at `011c9e6`, model Qwen3.8-Flash-Next IQ3_XXS, 262144-token context, MTP on, xhigh thinking. The only difference between the runs is speculative-decoding depth: `--spec 2` vs `--spec 3` (sequential runs, never at the same time).

**Report:** `bench/results/2026-10-22-community-mac-studio-m5-ultra-iq3-xxs-spec2-spec3/README.md` (per-run evidence in `spec2/` and `spec3/`: README, results.json, outputs.jsonl, config.json, provenance/; plus `needles-spec2/3.json`).

**Answers TEST_REQUESTS Mac row:** "A Mac with Apple silicon (macOS 15 or newer)" — currently `none yet`. This is, to my knowledge, the first Apple-silicon Mac standard-benchmark report: macOS 27.2, Mac Studio M5 Ultra 96 GB, real model, speed numbers, no failures.

**Headline results (medians):** short decode 133.8 → 142.3 tok/s (+6.4% for spec3); prompt-4k 2614.3 → 2579.7 (-1.3%); prompt-16k 2656.1 → 2571.7 (-3.2%); 16k decode 121.8 → 110.7 (-9.1%). Sanity 12/12 on both; needle recall 6/6 on both.

**Stages:** required suite complete on both runs (21/21 measured requests + warmup, performance + sanity). Optional concurrency stage: not run. Failed/skipped: none on either run.

**Nonconforming / limitations (see README):** single run per configuration (no run-to-run variance); the 4K/16K performance requests hit the 512-token cap while still reasoning so first-answer time is partly missing (protocol-cap property, not a failure); prompt-prefix reuse is template only; storage link speed / power / cold-model timings not measured; home/user/host redacted (grep-inspected clean). No engine changes in this PR.

Validation: numbers above are the README's medians, rebuilt from the committed results.json; artifact sha256 hashes are listed in the README.