Private AI on your Mac just got a free upgrade, and it runs on hardware you likely already have.
What's happening
A developer published Backburner this past week, and the concept is genuinely clever. Your iPhone 15 Pro or newer has a Neural Engine and GPU that sit mostly idle while your Mac handles AI work. Backburner connects the two devices over a 10 Gb/s USB-C cable and splits model layers between them: the Mac handles layers 1 to 40 of a 27-billion-parameter model, and the iPhone handles layers 41 to 64, pipelining work back and forth over the wire.
On an M4 Pro MacBook Pro with 24GB of RAM, running the Qwen3.8-27B model, the developer measured a 44% speedup on prompt processing at longer context lengths. The effective context window grew from roughly 64,000 tokens to 128,000. That second number matters. It is the difference between a model that forgets the beginning of a long document and one that can hold an entire grant application, a full email thread, or a 60-page contract in memory at once.
Backburner is MIT-licensed, based on a fork of llama.cpp, and currently in pre-release. One person built it. It is already catching attention fast.
For teams handling sensitive data (patient records, donor information, legal files) who want real AI capability without sending anything to an outside API, this is exactly the kind of setup worth knowing about.
Try this this week
You will need an Apple Silicon Mac (M1 or newer), an iPhone 15 Pro or newer, and a 10 Gb/s USB-C cable. The stock cable in the iPhone box is too slow; a capable one runs about $20.
- Run the one-command installer from the README at github.com/StayLameBro/backburner
- Install the companion iOS app using AltStore or Xcode (App Store submission is pending)
- Download the Qwen3.8-27B model through the script in the repo (about 24 GB)
- Plug in the fast cable, launch both apps, and run a prompt against a real document your team uses regularly
The test worth doing: hand it a long client email thread or a multi-page report and ask it to summarize or pull out action items. Compare the quality against what your current 7B or 8B model does with the same input.
The bigger picture
This is the smallest-working-thing idea applied to hardware. Instead of buying a 64GB Mac or spinning up a GPU server, you connect two devices you already own and close most of the gap. The quality difference between a 27B model and an 8B one is real and noticeable on everyday tasks like summarizing, drafting, and analyzing documents. That kind of upgrade used to cost thousands of dollars in hardware. Now it costs a $20 cable.