EXO (EXO Labs)

EXO runs supported AI models on local hardware and can distribute their work across multiple devices. It is intended for developers, researchers and users experimenting with local inference, model serving and clusters whose combined memory can accommodate a model.

Inference and integrations

• Discover devices on a local network and inspect the cluster in a dashboard.
• Shard supported models according to device resources and topology.
• Serve compatible OpenAI Chat/Responses, Claude and Ollama API interfaces.
• Download supported model cards or configure custom Hugging Face models.
• Use local model storage, optional read-only model directories and an offline mode.
• Configure a namespace to avoid accidentally joining another cluster.

Compatibility varies by model, backend and API; an API-compatible endpoint does not imply every provider feature is implemented. Version 1.0.71 adds Kimi K2.6 multimodality, provider sampling defaults and fixes for M5 vision inference and RDMA output.

Installation and hardware

Download the official 1.0.71 DMG and copy the app to Applications, or use brew install --cask exo. The exact official DMG is 412,972,413 bytes, verified by HEAD. Model downloads, caches and logs consume additional disk space and may be many gigabytes.

The version-tagged 1.0.71 README explicitly requires macOS Tahoe 26.2 or later for the packaged Mac app. This takes precedence over the catalog's older macOS 15 condition. The catalog targets Apple Silicon; the project uses MLX/Metal and documents Apple Silicon monitoring. The checked sources do not establish Intel support for this packaged build or a universal numeric minimum RAM. Required unified memory and storage depend on the chosen model, quantization and participating nodes; clustering is not a guarantee that an arbitrary model will fit.

Ordinary clustering needs a reachable local network. Optional RDMA requires macOS 26.2 and Thunderbolt 5 hardware and additional system configuration; it is not available on every Apple Silicon Mac. Source builds separately require Xcode's Metal toolchain, uv, Node and Rust, with pinned macmon guidance for M5. These development tools are not described as prerequisites for installing the packaged app.

Permissions, network and privacy

The app asks to modify system settings and install a Network profile. Review these elevated changes before granting permission. Downloading models or updates needs internet access and some gated model repositories require a provider account or acceptance of their model terms. Offline mode uses already downloaded local models; local cluster networking remains necessary for multiple nodes.

Local inference avoids sending prompts to a remote inference provider, but clients, integrations and model downloads can still make network requests. Treat the API and cluster as services on a trusted network, protect prompts and model files, and do not expose endpoints to an untrusted network without independently configured access controls. A custom namespace prevents accidental cluster joining; it is not evidence of authentication or encryption. Remote model code is disabled by default; enable trust_remote_code only for code you have reviewed and trust.

License and maintenance

EXO is Apache-2.0 open-source software. Model weights and integrations retain their own licenses, acceptable-use terms and possible charges. There is no required EXO cloud subscription described for local inference. Performance, output accuracy and stability depend on the model and hardware; verify important answers and monitor memory pressure, heat and storage.

Use the app's Advanced → Uninstall command to remove its system components. Simply deleting the app can leave the network setup LaunchDaemon, scripts, logs or network location behind; the official README documents cleanup and notes that Login Items may require manual removal.

Sources: Version 1.0.71 README and requirements, Exact release notes, Official download, Project.