I made my iPhone a second GPU for my 24 GB MacBook: Qwen 3.8 27B prefills 29–44% faster & my holds part of the CTX window.

**DISCLAIMER** THE PREFILLING TPS SHOWN ON THE PHONE IS COMPUTED ONLY FOR THE LAYERS IT HOLDS. ALREADY FIXING IT TO SHOW END-TO-END PREFILL RATE. NUMBERS BELOW ARE ACCURATE FOR E2E PREFILL RATE.

Every file or tool result my agent reads on a 24 GB M4 Pro MacBook is a wait, and 64k of 8-bit context is all that fits next to Qwen 3.8 27B (IQ4_XS), even with the wired limit raised to 20480. An iPhone 17 Pro Max was sitting in my pocket, so I figured what can I do to make use of this extra silicon.

Turns out a 10 Gb/s USB-C cable & some software is all you need. The Mac runs layers 1–40 of each 256-token batch and streams the activations to the phone. The phone runs layers 41–64 on its GPU while the Mac starts the next batch. The A19 Pro's GPU has matrix units (Metal 4 tensor ops), and they make the phone's half 2.4x faster than the same phone without them.

Same build, phone off vs. on, prefilling a 2,000-token file into a saved agent session:

• 8k context: Mac alone 132 tok/s → Mac + iPhone 177 tok/s (+35%) (measured two days earlier, same bench)
• 16k context: Mac alone 109 tok/s → Mac + iPhone 157 tok/s (+44%)
• 32k context: Mac alone 101 tok/s → Mac + iPhone 130 tok/s (+29%)
• 48k context: Mac alone 87 tok/s → Mac + iPhone 113 tok/s (+30%)

A fresh 27k-token agent session, cold: 245 s on stock llama.cpp, 228 s on my fork with the Mac alone, and 168 s with the phone.

Past 64k the phone switches jobs. The oldest KV pages move to the phone and the Mac runs all 64 layers. For every attention layer, the phone computes attention over the old keys on its GPU, and the Mac merges that with its own part. While writing, the phone's Neural Engine takes part of that work too: each 16k-key page of old context is compiled into a Neural Engine model with the keys as its weights. At 140k that took writing from 279 to 176 ms per token compared with the phone's GPU alone.

The server allocates 196k–229k of 8-bit context based on the phone's free memory; that's up to ~5.7 GB of KV cache living on the phone instead of the Mac, so the Mac's memory use stops growing at 64k. I've tested a growing session to 128k at 8-bit, with 3/3 planted facts recalled. Separately, at 140k in 4-bit, the run passed the gate with greedy output matching the Mac-only run for 32 generated tokens.

What it doesn't do: speed up writing below 64k. That's the Mac's job. My fork's kernels (SME2 on the M4 CPU and Metal fusions) plus DFlash2 speculative decoding take it from 11.3 tok/s on stock llama.cpp to 25 tok/s at about 30k context with medium thinking, phone or not. SME2 also adds up to 29% to prefill on the Mac alone. Past 64k the phone does share the writing (attention over the old keys), and without it the Mac would have to drop to 4-bit context to reach 128k. In real use I have seen upwards of 30 TPS at lower context.

The phone joins prefills over about 512 tokens. In one real session, that was 7 of 36 requests, but about 83% of the tokens read. Past 64k it holds the context and does the old-key attention, but it stops running layers 41–64 there for now; doing both is next. One request at a time.

I'm curious what this setup could do with newer model architectures. DeepSeek V4.1-Flash reports 890 bytes per token for its global KV cache and adds n-gram embedding tables (Engram). Qwen3.8-Flash-Next, the Qwen 4 architecture preview, has an n-gram lookup table too. Those aren't features of the 27B model I tested, and I haven't benchmarked either architecture here. The real gold is within the newer phones and models working together. With the A20 Pro in the iPhone 18 Pro Max, I bet there is a lot more for me to push.

Code, setup and bench scripts: https://github.com/StayLameBro/backburner

Still a lot of work to do but I built this with Opus 5.5. Happy to answer anything.