Qwen3.8 27B MLX 4-bit; 0.7.0; client-observed single-stream decode

Developer-reported measurement by Rapid-MLX, measured 2026-10-12, read 2026-10-13; not rerun locally. Primary comparison: https://rapidmlx.com/compare . Exact Qwen3.8-27B summary: https://rapidmlx.com/blog/assets/compare-2026-10/27b/summary.json . Versions: https://rapidmlx.com/blog/assets/compare-2026-10/27b/raw/versions.json . Harness: https://rapidmlx.com/blog/assets/compare-2026-10/bench_compare.py . Per-request inventory: https://rapidmlx.com/blog/assets/compare-2026-10/27b/raw/manifest.json . One Mac Studio, Apple M5 Max, 36 GB installed unified memory. CPU/GPU core counts, macOS, storage, power mode, thermals, clocks, swap, detailed placement and measured RAM/VRAM are unstated in the inspected October metadata. Installed unified memory is not dedicated VRAM. Qwen3.8-27B dense, MLX 4-bit. Engine/configuration: oMLX 0.7.0. Documented SSD cache directory enabled; image input supported. Speculative-decoding state is not established by the inspected metadata. The comparison says the engines used identical weights. Verified upstream reference: https://huggingface.co/Qwen/Qwen3.8-27B . This is NOT the tested MLX artifact. Current alias documentation points qwen3.8-27b-4bit to https://huggingface.co/rapid-mlx/Qwen3.8-27B-4bit-MTP-MLX , but the inspected run metadata does not bind that repository, its revision, shard hashes or the exact drafter to these runs. Do not present the current alias mapping as a verified historical download. Quantization group size/recipe and tested revision remain unknown. Headline is 25.262689120415786 tokens/s, the published summary's single-stream decode figure. Its numerator is the engine's completion-token count minus one; its denominator is client-observed elapsed time between first and last content-bearing SSE chunks. Reasoning output is counted when reported by the engine. This is a streaming decode-phase rate, not engine-native timing, end-to-end throughput or prompt processing. Chunk grouping and transport can affect comparisons. The harness computes a median of three requests for this scenario, and summary decode_runs is 3; the comparison prose also calls this a single run per engine. Raw request files were not obtained, so this discrepancy and sample spread are unresolved. The harness requests an essay about bicycle history, temperature 0, maximum 512 generated tokens. Actual decode prompt and completion counts are unknown here; 512 is the requested cap. Configured context capacity is unknown. The harness normally warms up with a separate 16-token-cap greeting, but the October invocation flags and effective cache state are unverified. Do not inherit the September machine, artifact revisions or thinking flags. Separate four-stream result: 75.30712308042115 tokens/s. That scenario sums engine completion counts over wall time for all four requests, including their startup/prefill and waiting, then takes the median of two repetitions. It is NOT single-user decode and must not be placed in the headline generation field. Its per-request cap is 256; realized counts are unverified. Separate cold TTFT probes: 2030 reported prompt tokens with 2.6151716250169557 seconds, and 15533 with 19.888424208998913 seconds. These counts are summary metadata for other requests, not the decode prompt or configured capacity. TTFT begins at request send and ends at the first content/reasoning/tool-bearing chunk. Native prefill throughput is unavailable: leave PP blank and do not divide tokens by TTFT. Flash Attention, KV precision and exact batch configuration are unknown. Source warns differences below roughly 5% can be run variation. Findings are source-reported developer measurements with the above reproducibility limits.