Qwen 3.8 27B Just Hit 200 Token/s on Mac M5 (Lithos Metal)

LithosAI says Qwen 3.8 27B runs at more than 200 tokens a second on an M5 Max, and its own measured run five days earlier works out to about 80. lithos-metal is a new open-source Metal engine for Apple silicon. It uses DSpark speculative decoding: a 1.1 GB drafter guesses seven tokens, and the big model checks all of them in one read of its weights. That's how a 40-core M5 Max, which tops out near 38 tokens a second on the 16 GB 4-bit Qwen 3.8 27B file (614 GB/s ÷ 16 GB), can go past its own memory ceiling. The README's 168 tokens a second is a projection that assumes six of those seven guess