EXO — Distributed AI Platform

Run large AI models across your existing devices. No cloud, no new hardware, no cost.

Quick Start (Mac only)

git clone --recurse-submodules https://github.com/YOUR_USERNAME/YOUR_REPO
cd YOUR_REPO
chmod +x setup.sh start.sh
./setup.sh      # one-time setup (~5 min)
./start.sh      # run every time

Open http://localhost:5500 in your browser.

What it does

• Splits AI models across all devices on your WiFi network
• Each device holds a shard of the model in RAM
• Run 20B+ parameter models that wouldn't fit on a single machine
• Tracks CO₂, water, and energy saved vs cloud GPU equivalents

Adding more devices

On any other Mac on the same WiFi:

git clone --recurse-submodules https://github.com/YOUR_USERNAME/YOUR_REPO
cd YOUR_REPO
./setup.sh
./start.sh

It joins automatically. No configuration needed.

Requirements

• macOS (Apple Silicon or Intel)
• Python 3.13 (installed automatically)
• Rust (installed automatically)
• 8GB+ RAM recommended
• Same WiFi network as other nodes

Models included

Model	Size	Status
Llama 3.2 1B 4bit	696MB	Downloaded on setup
Qwen3 0.6B 8bit	666MB	Available
Llama 3.2 3B 4bit	1.7GB	Download via UI
Llama 3.1 8B 4bit	4.3GB	Download via UI

Architecture

[Your Mac]          [Friend's Mac]       [Any Mac on WiFi]
exo + frontend  +   exo worker       +   exo worker
     └──────────────────────────────────────┘
              Auto-discovered via mDNS
                    │
            http://localhost:5500

Environmental impact

Every token processed locally avoids:

• ~0.18g CO₂ per 1000 tokens (vs AWS GPU)
• ~2.4mL water per 1000 tokens (vs data center cooling)
• ~0.38Wh energy per 1000 tokens

Stack

• exo — distributed inference engine
• FastAPI — CORS proxy
• Vanilla JS — frontend