- What an inference cluster is and when it's worth it.
- Using exo to distribute a model across multiple machines.
- Connecting computers over the local network (Ethernet or Thunderbolt).
Key concepts
exo: the home cluster
exo (from exolabs) is an open source project that automatically connects the computers on your network and distributes the model among them. In 2026 it is a mature tool: it detects machines, balances load based on each one's power, and takes advantage of fast connections like Thunderbolt.
- Two or more computers (Macs with M chips and/or PCs). The more you have and the more powerful they are, the larger the models.
- All on the same local network. Ideally connect them via Ethernet cable (or Thunderbolt between Macs): much faster and more stable than Wi-Fi.
- Claude Code on one of them to guide you through installing exo on each machine.
Step by step (overview)
You'll see the exact setup in exo's documentation, but the idea is this:
- Connect all computers to the same network (cable is better).
- Install exo on each machine. Ask Claude Code on each one: "help me install exo and verify it starts up."
- Start exo on all of them. They discover each other automatically and form the cluster.
- From any of them, launch a large model: exo distributes it across the machines and exposes a common access point for your applications.
If something goes wrong
- They can't see each other — confirm they're on the same network and that no firewall is blocking exo.
- It's very slow — switch from Wi-Fi to cable; check that no machine is overloaded by another task.
- One machine drags down performance — exo distributes based on capacity, but a very weak node can hurt things; try it with and without it.
Practice challenge
Measure the difference: run the largest model that fits on a single machine, then a larger one with the cluster. Compare which models you can use in each case. You'll have built your own AI infrastructure at home.