开发者联用 MacBook Pro 与 iPhone 运行 Qwen，称预填充提速 44%

Summary

IT Home reported on October 4, 2026: A developer used the open-source backburner project to split Qwen3.8-27B layers between a MacBook Pro and an iPhone 17 Pro Max. In a reported 16K-context test, prefill throughput rose from 109 to 157 tokens per second, about 44%, using the Mac GPU and the phone’s neural engine. The result applies to one hardware, quantization and software setup; prefill is not full generation speed, energy use or stability. Information reflects the publication date and should be checked against original documents and later independent assessments.