Conversation data remains local
A supported Minirun chat assembles the prompt and executes the model on the device. Conversation history is stored in the app's private data. Minirun does not require a cloud inference endpoint for the prompt or generated response.
When the network is used
The app contacts Hugging Face to browse current Minirun publications, resolve a repository's immutable file tree, and download model files. A catalog refresh is separate from opening a verified local model or sending a message.
Remote repository metadata can be cached so the product remains understandable when the catalog is temporarily unavailable. Local model truth still comes from registered storage and verification evidence, not from a cloud-installed badge.
Folder permissions
macOS and iOS require you to choose external folders through the system picker. Minirun saves only the authority required to reopen registered model locations and per-download destinations. Removing a storage location forgets that saved access without deleting its files.
Verification before execution
Execution requires matching full-verification evidence for the publication, container metadata, payload identities, and local filesystem objects. These checks protect the meaning of the bytes consumed by the runtime; they do not replace normal operating-system, account, or storage security.