TURN FOUR PHONES INTO ONE AI COMPUTER

A distributed AI setup that turns recent Android and iOS devices into a single local inference cluster.

by orailnoor

AISoftware

difficulty
●●●○○
time
an evening
license
license not specified
repo
repo
1
Jump to section

COMPAREE VERDICT

This is a practical demonstration of distributed inference: you install PrivateLM on multiple phones, designate one Android device as master, scan QR codes to add workers, and the app splits a 26 billion parameter model (Google Gemma 4 26B) across devices according to their memory. The iPhone handled 38%, the iPad 32%, a Galaxy S24 21%, and a Pixel 9%. Each device runs part of the model's layers and passes results to the next over Wi-Fi, so you get one model running across four devices with no cloud. The server tab exposes an OpenAI-compatible API on your home network, so a laptop can query it or use it for coding in VS Code. The catch: it is slow at 4 to 5 tokens per second, every additional phone adds Wi-Fi latency, the master must be Android, and the demo used recent devices, not old phones from a drawer. If you have a few modern phones sitting idle and want to experiment with local LLM distribution without buying new hardware, this is a weekend well spent. The most likely frustration is Wi-Fi performance — a congested or slow network will make the already-slow inference even slower.

GOOD TO KNOW

  • —Free and open source app available on GitHub (orailnoor/cross-platform-llm-client).
  • —Master device must be Android; workers can be Android or iOS.
  • —No hardware build — this is a software setup using existing phones and tablets.
  • —Video walkthrough shows the full setup process and performance.
  • —Runs entirely local with no cloud dependency; OpenAI-compatible API server mode included.

Can I build this?

Printnothing required
Buynothing new if you already have 2–4 recent phones or tablets (Android and/or iOS)
Toolsstable Wi-Fi network, GitHub account to download the app
Skillscomfortable installing apps from GitHub, basic understanding of local AI models and network APIs
Timean evening to install, configure QR pairing, and test inference; longer if troubleshooting Wi-Fi or model download
Cost$0 if using existing devices; otherwise cost of phones (recent models recommended)
SafetyNo safety risks beyond ordinary device use. Runs entirely offline, no mains voltage, no hardware modification.

Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.

Videos

Running Gemma 4 26B across four phones with PrivateLM

Full walkthrough by the developer, showing setup and real performance.

More builds like this

All projects

Gallery

Start here

Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.

  1. 1.Check device compatibility(Master must be Android. Workers can be Android or iOS. Recent phones recommended (demo used iPhone, iPad, Galaxy S24, Pixel).)
  2. 2.Install PrivateLM on all devices(Follow installation instructions for each platform.)
  3. 3.Set up phone cluster(Designate the strongest Android phone as master. Use QR code pairing to add each worker device. Keep all devices on the same Wi-Fi network.)
  4. 4.Load the model(The app will download and split Google Gemma 4 26B across devices by memory. Wait for the model to distribute.)
  5. 5.Test inference(Run a test prompt to verify all devices are communicating. Expect 4–5 tokens per second.)
  6. 6.Enable API server (optional)(Turn on the server tab to expose an OpenAI-compatible API on your home network. Use it from a laptop or in VS Code.)

Resources

Documentation, files and community threads for this build — we link straight to the original sources and never rehost the creator’s files.

KNOWN ISSUES

  • Assuming old phones will work — the demo used recent models (iPhone, iPad, Galaxy S24, Pixel). Older devices may lack the memory or performance.
  • Choosing an iOS device as master — the master must be Android. An iPhone or iPad can only be a worker.
  • Running on congested Wi-Fi — every additional phone adds a Wi-Fi hop, and a slow or busy network will make the already-slow inference even slower. Use a dedicated or quiet network.
  • Expecting fast inference — 4 to 5 tokens per second is what the demo showed. This is slow compared to cloud APIs or dedicated hardware.
  • Running out of storage or patience on the model download: a 26B model is large, so keep every device charged and on Wi-Fi while it loads.
  • Not testing the API before integrating — verify the OpenAI-compatible server works with a simple curl or Postman request before trying it in VS Code or another tool.

Can I use old phones from a drawer?

The demo used recent devices (iPhone, iPad, Galaxy S24, Pixel). Older phones may not have enough memory to hold their share of the model or may be too slow to keep up. You can try, but expect worse performance or failures.

Does the master have to be Android?

Yes. The master device must be Android. Workers can be Android or iOS.

How slow is 4 to 5 tokens per second?

Slow enough to be annoying for interactive chat. This is for experimentation, not production use.

Can I use this on a public Wi-Fi network?

Not recommended. All devices need to be on the same network, and public Wi-Fi often blocks device-to-device communication. Use a home or private network.

What happens if one phone disconnects?

Inference will fail because the model is split across devices. All phones must stay connected and on the same network for the duration of the session.

Can I use this for production or commercial applications?

This is an experimental setup. The slow inference speed, Wi-Fi dependency, and need to keep multiple devices running make it impractical for production. It is best suited for learning or demonstration.

Community builds

No community builds yet — be the first, we feature the best ones.

Discussion1

FROM THE COMPAREE TEAM

The demo ran Gemma 4 26B at 4–5 tokens per second across an iPhone, iPad, Galaxy S24, and Pixel. What would you use a local, multi-phone AI cluster for that makes the slow speed worth it?

CompareeTEAM13d ago

The video is a clean walkthrough showing the full setup, QR pairing, load distribution by percentage, and the OpenAI API server working in VS Code. The master-must-be-Android requirement is easy to miss until you start, and Wi-Fi quality matters more than you would think — each phone is a network hop. The most useful part is the API server: once running, the cluster acts like a local OpenAI endpoint, so any tool that speaks that protocol can use it. Slow, but genuinely local.

orailnoor

Developer who built PrivateLM to demonstrate distributed local AI inference across multiple mobile devices, showing that model splitting can turn consumer hardware into a collaborative compute cluster.

Star the project on GitHub

DISCLAIMER

  • Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
  • Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
  • Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.