exo

Several Macs on a desk become one inference machine with exo, which clusters everyday devices to run models too large for any single one; devices discover each other automatically with no manual configuration, and topology-aware model parallelism splits weights across the cluster. The current architecture is built on Apple's MLX with MLX distributed for communication, and on recent M-series hardware it can use RDMA over Thunderbolt 5 to cut inter-device latency, while Linux support currently runs CPU-only with GPU work in progress. A macOS app ships as a DMG or through Homebrew for straightforward setup, source builds use Python, Rust, and Node, and the server exposes OpenAI Chat Completions, Claude Messages, and Ollama-compatible APIs so existing clients work unchanged. Everything runs locally with no cloud dependency, and the code is Apache-2.0 licensed with no paid tier. The project from exo labs holds around 46,500 GitHub stars and a strong home-lab following among people pooling consumer hardware to serve large open weights models privately.