Qwen3.8 27B

User benchmarks Qwen3.8-27B across 15 Mac inference setups on an M5 Pro 64GB, with the fastest decode at 78.6 tok/s in Inco Splash 1.6.0 using a Jundot oQ4e pack with DFlash2 speculation. Each run sent one OpenCode-style coding-agent request of about 88K input tokens with tool schemas, a 2,048-token output limit, reasoning effort medium, temperature 0 and seed 17 where supported; engines ran sequentially and the Splash row is the median of three cold runs. Decode spans 8.8 tok/s (Jan 0.8.8) to 78.6 tok/s and prefill 99.9 to 614.1 tok/s across the table; the user notes model packs, quantizations, speculative decoders and context allocations differ, so it is not a controlled engine-only comparison.