Uzu inference engine adds speculative decoding, 3x faster than llama.cpp on M5

Mirai Labs shipped speculative decoding in its local inference engine Uzu, debuting with Qwen3.6 27B and reaching 105 output tok/s on an M5 Max with 128GB — about 3x faster than llama.cpp.

2026-09-04 ~ 2026-09-04 · 3 related posts

• Mirai's Uzu engine hits 105 tok/s with Qwen 27B on M5 Max, 2.9x faster than MLX speculative decoding — awnihannun · 2026-09-04
• Uzu lab ships speculative decoding implementation, launching with Qwen3.6 27B support — awnihannun · 2026-09-04
• Mirai's uzu engine brings speculative decoding to Apple M5, hitting 105 tok/s on Qwen3.6 27B — TheMoonMidas · 2026-09-04