Your iPhone is the extra GPU your Mac LLM was missing

What it does

Backburner is a llama.cpp fork that splits a 27B model (Qwen3.8-27B) across a Mac and a tethered iPhone. During prefill, the Mac runs layers 1-40 and streams the residual to the phone, which runs layers 41-64 on its GPU while the Mac starts the next batch. Past 64k context, the phone changes jobs: it holds the oldest KV pages and computes attention over them, so a 24 GB Mac can carry 8-bit context out to 196k-229k tokens instead of dropping to 4-bit at 64k.

The interesting bit

The phone’s Neural Engine gets used in a genuinely odd way: since old keys never change, each 16k-key page is compiled into an ANE model with the keys baked in as weights. That plus the GPU cuts old-key attention from 279 to 176 ms per token at 140k. Also notable: greedy output is token-identical with and without the phone — this is a speed and memory play, not an accuracy tradeoff.

Key highlights

• 29-44% faster prefill at 16k-48k context on a 24 GB M4 Pro + iPhone 17 Pro Max
• Context sized from the phone’s free memory at startup; tested end-to-end to 128k at 8-bit
• Mac-side wins too: SME2 CPU co-compute, Metal fusions, DFlash2 speculative decoding (25 vs 11.3 tok/s decode vs stock llama.cpp)
• SSD prompt cache restores a 27k-token session start in 0.3-5 s
• Community-verified on 8 GB Macs, iPads, and older iPhones (with varying results)

Caveats

• Reads under ~512 tokens stay entirely on the Mac; small agent steps don’t benefit
• iPhone 16 Pro Max reportedly runs hot over long prefills; non-Pro iPhones’ USB 2 ports are too slow for split prefill
• Some reported setups are still being debugged (128k on iPhone 18 Pro Max, occasional unexplained iPad failures)

Verdict

If you run local models on Apple hardware and own a Pro iPhone, this is one of the more creative uses of that cable in your drawer. If you’re on Linux/Android or expect it to fix decode speed below 64k, wait for the roadmap.