Qwen3.8 27B Generates Half as Fast as Its Predecessor on a Mac Studio but Finishes in About the Same Wall Clock

TerminalBytes / Hacker News·medium signalXBlueskyLinkedInCopy link

TerminalBytes benchmarked Qwen3.8 27B on an M3 Ultra Mac Studio with 256GB on August 27, measuring about 14 tokens/sec at Q4_K_M (17GB file) against roughly 28.6 tokens/sec for Qwen3.6 27B on the same machine. The catch that makes it usable is that the newer model answers in roughly a third fewer tokens, so wall-clock time is comparable; a 1-bit quant at 6.7GB hits 27 tokens/sec and fits in 16GB. Anyone trying this needs a llama.cpp build from the last few weeks, since older ones fail with 'unknown model architecture: qwen35'.