exo
Turn your everyday devices into one AI cluster to run frontier models.
About
exo connects the devices you already own — Macs, PCs, phones — into a single AI cluster, so you can run models that would never fit on one machine. Devices discover each other automatically, and exo splits the model across them with topology-aware auto-parallelism, including day-0 support for RDMA over Thunderbolt for very low inter-device latency.
It's compatible with the OpenAI Chat Completions, Claude Messages, OpenAI Responses and Ollama APIs, so existing tools and clients work against it unchanged. Maintained by exo labs, Apache-2.0 licensed.
What it can do
- Multi-device distributed inference
- Automatic device discovery
- Topology-aware model splitting
- OpenAI/Claude/Ollama API compatibility
- Run models larger than one device
Setup
- Install exo on each device you want in the cluster (see the repo for per-OS instructions).
- Start exo on each device — they discover each other automatically, no manual config.
- Download or point exo at the model you want to run.
- Send requests to its OpenAI-compatible API endpoint from any client.
Prerequisites
Two or more devices on the same network (Mac, Linux, Windows); Thunderbolt for the fastest interconnect.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 47.8k stars — huge interest in device-cluster inference; pushed 2026-10-12 — active. Manual review: OpenAI/Claude/Ollama API compatibility verified; multi-device tensor parallelism is ambitious — real-world throughput varies with interconnect. License: Apache-2.0.
★ 47,849 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · 47.8k stars — huge interest in device-cluster inference
- · pushed 2026-10-12 — active
- · manual review: OpenAI/Claude/Ollama API compatibility verified; multi-device tensor parallelism is ambitious — real-world throughput varies with interconnect
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.