iPhone as a second GPU speeds up Qwen 3.8 27B prefill by 29–44%

On October 3, 2026, a developer posted on Reddit's LocalLLaMA community about using an iPhone 17 Pro Max as a second GPU alongside a 24GB M4 Pro MacBook. The phone carries part of the context window for Qwen 3.8 27B at IQ4_XS quantization, and according to the developer, end-to-end prefill speed improved by 29–44%.

The backdrop, as described in the post, is memory pressure on the MacBook. Every file or tool result the developer's agent reads involves a wait, and even with the wired limit raised to 20480, only 64k of 8-bit context fits next to Qwen 3.8 27B (IQ4_XS).

The post carries a disclaimer: the prefill tokens per second shown on the phone is computed only for the layers that device holds, so it does not represent overall throughput. The accurate figure, the developer says, is the end-to-end prefill rate, and the on-device display is already being fixed to report it. The numbers given in the post, the developer stresses, are accurate for end-to-end prefill rate. As of this report, the work stands at the point where the developer has published the figures and flagged that the display metric still needs correcting.