iPhone 17 Pro Max boosts 27B model prefill on M4 Pro MacBook Pro by up to 44%

A user connected an iPhone 17 Pro Max to an M4 Pro MacBook Pro over USB-C and, using open-source software called backburner, split work on the Qwen3.8-27B model: the Mac runs layers 1–40 while the iPhone runs layers 41–64 on its A19 Pro GPU. Prefill at 16K context rose from 109 to 157 tokens/s (+44%), at 8K from 132 to 177 (+35%) and at 32K from 101 to 130 (+29%).

• Prefill at 16K context rose from 109 to 157 tokens/s, a 44% gain
• At 8K it went from 132 to 177 tokens/s (+35%), at 32K from 101 to 130 (+29%)
• Mac handles layers 1–40, iPhone 17 Pro Max runs layers 41–64 on A19 Pro GPU
• Speedup applies only to prefill; text generation stays on the Mac