Skip to main content
Start with Quick Start to choose a device and a model, then follow the matching platform guide to build and install Edge0. Building from source requires a terminal and the platform’s build tools.
For a first try, 8B is recommended: the download size and resource footprint are both smaller. When your device has enough resources, try 35B and compare the two on your typical tasks.For specifications, device requirements, and a speed reference, see Models: 8B vs 35B.
No. A network connection is needed to download the source, dependencies, and models; once preparation is complete, responses are generated locally without using a cloud model service.Downloading new models, opening external links, or sending feedback still uses the network. See Privacy & Data Handling for details.
The app source and the models are provided separately, and cloning the repository does not download any model. Obtain models through the official sources in Download & First Run, and prepare them following the platform guide.
No. The model format differs by platform: macOS uses the released model files directly; Windows and Android need conversion to GGUF; iOS 35B needs a file-layout conversion first. Prepare the files separately following each platform guide, and use matching files from the same version.
After the download finishes, the model still has to be loaded. On Windows, the model is converted locally first; on iOS, the model must be bundled into the app at build time; on Android, the model must be imported and the list refreshed first. For how to handle this, see Troubleshooting.
It means the model was not bundled into the currently installed app. Place the model into the corresponding directory under Models/, then rebuild and reinstall the app.
The first model load and the first turn of a new topic have to read model data from storage first; a longer input also needs more processing time. Later replies are usually faster.If you also see error messages, crashes, or a long period of no progress, see Troubleshooting.
35B refers to the model’s parameter scale, not a memory size. Quantization reduces the model file size, and an MoE model uses only part of its parameters per step, reading the rest from storage on demand. Choose a device according to the Device Requirements.
No. These figures are the measured resident memory values for running 8B and 35B, respectively, on a MacBook Air M3; the system and other apps need additional memory.
The installer contains only the desktop app; at runtime it also needs the local engine and the model conversion tools. Prepare these on the target computer as in the Windows Guide.
The current macOS app uses ad-hoc signing and the Windows installer is unsigned, so the system may show a security prompt. Verify the source first, then follow Security & Installation Warnings.
The current models are preview releases, and their coverage and quality are still improving. The model cards note that tool use, multi-step planning, and long-horizon autonomous tasks are not yet well optimized. When using them for important purposes, verify the model’s output yourself.
Edge0’s code and both model tiers are under the Apache-2.0 license, which permits commercial use provided you comply with the terms. The base models and third-party components are subject to their own terms; see License & Model Terms.
When reporting, please provide:
Try to reproduce the problem with a short input that contains no private information. For how to obtain logs, see Troubleshooting.GitHub Issues is a public channel; do not upload private conversations, access tokens, your entire home directory, or model files. Remove sensitive content from email attachments as well.