Qwen3.8-27B

Same Mac. Speculative decoding tuned.

+45%faster.

Measure yours!

asiai bench

$ asiai bench -m qwen3.8-27b

draft-mtp 4 31.8 tok/s

no speculation 21.9 tok/s

M5 Max · llama.cpp b10270 · Q5_K_XL · n=3

10 engines GPU observability Open Source Zero dependencies Leaderboard MCP Claw-ready Agent-ready