Oct 13, 2026
Summary
Source-reported developer measurement by Rapid-MLX, measured 2026-10-12 and read 2026-10-13; not rerun locally. Primary source: https://rapidmlx.com/compare . It links the common timing harness: https://rapidmlx.com/blog/assets/compare-2026-10/bench_compare.py . The Qwen3.5-9B October summary, versions and per-request files were not obtained, so this report uses the article's model-specific table/measurement notes and retains rounded rates.
Hardware: ONE Mac Studio, Apple M5 Max, 36 GB installed unified memory. CPU/GPU core counts, OS, storage, power/thermal state, swap, detailed placement and measured memory are unknown. Installed unified memory is not dedicated VRAM.
Model: Qwen3.5-9B, MLX 4-bit. Runtime: LM Studio llmster 0.0.25; MLX runtime 1.11.0. Headless LM Studio with MLX runtime. Speculative-decoding state is not established. The source says the comparison engines use the same weights. Upstream model verified at https://huggingface.co/Qwen/Qwen3.5-9B ; this is an upstream reference, not the tested MLX download. Tested converter/repository, revision, checksum, shards, precise quantization recipe and drafter artifact are unverified. Do not inherit the September artifact revisions or map today's alias to the historical download.
Headline: 81.1 tokens/s, the article's rounded single-stream decode rate. The shared harness defines decode as engine-reported completion tokens minus one divided by the interval between the first and last content-bearing SSE chunks observed at the client. Reasoning tokens are included when returned in usage. This is client-observed streaming decode-phase throughput, not engine-native timing, wall-time completion throughput or prompt processing; stream chunking may affect comparisons.
The published scenario-a harness asks for a bicycle-history essay at temperature 0, a 512-token output cap, and computes a median of three requests. The October prose describes a single run per engine. With this model's raw summary unavailable, the realized repetition count, sample spread, actual prompt and completion counts, invocation flags and warm-up/cache state are unverified. The script's normal separate greeting warm-up has a 16-token cap but can be disabled. A cap is not an actual output count, and the common script alone does not prove every invocation flag. Configured context capacity, KV precision and Flash Attention are unknown.
Separate four-stream result: 230.0 tokens/s. The common scenario-d harness sums engine completion counts over the elapsed wall time of all four requests, which includes prompt processing/startup and waiting; its standard protocol takes the median of two repetitions and asks for at most 256 tokens per request. This is an aggregate wall-throughput result, not single-user decode, and stays in Notes rather than headline TG. Actual output counts and repetitions are not raw-validated here.
The separate approximately 2K and 16K cold-prompt TTFTs are 2.780 s and 21.743 s. Their actual token counts are not verified. The main comparison table says 81.1 tokens/s, while the homepage says 81.0. Retain the main table value as rounded source-reported data, with this discrepancy disclosed; no raw adjudication was possible. TTFT is request send to the first content/reasoning/tool-bearing chunk; it must not be converted to native prefill by dividing prompt tokens by this time. No native PP is reported, so leave PP blank. Approximate probe sizes and the separate roughly 23K-token agent-session prefix are not decode prompt length or configured context capacity. The source warns that differences under about 5% may be run variation. Treat as a developer comparison with limited reproduction evidence, not independent validation.
Additional, separate October agent-replay latency observations from the same model/runtime table (seconds): initial cold turn 1 32.694; turns 2-10 median 0.613; total of ten turn timings 37.894. After server restart: turn 1 32.678; turns 2-10 median 0.611; ten-turn total 37.842. The article describes per-turn time to first token; if no earlier streamed token is emitted, the harness can fall back to end-of-stream time. Per-request fallback flags are unavailable here, so these are reported agent-turn latencies with that caveat, not verified pure-TTFT values, load-plus-prompt estimates, or the single-stream decode workload. The session prefix is approximately 23K tokens; do not use it as configured capacity.