Qwen3.8 125B (6B active) Flash-Next Uncensored

User reports Qwen3.8-Flash-Next Uncensored at a median 116.6 t/s decode over 363 real coding-agent requests on an M5 Max 128GB. Setup is Inco Splash 1.4.1 on Metal behind an oMLX proxy, Q5_K_S GGUF with int8 KV cache, DFlash2 speculative decoding on, single sequential request, all weights resident and the 51B n-gram table served from SSD cache. Median decode by prompt length was 124.3 t/s under 32K, 95.8 t/s at 32-64K, 104.6 t/s at 64-128K and 97.9 t/s at 128K+, with p25 100.9 and max 205.5 t/s. Median stream-visible TTFT was 0.87 s (p90 3.56 s) with a warm prefix cache; cold long prompts reached up to 68 s. Prefill rate was not isolated because prefix reuse dominated.