Local AI model storage on Mac: downloads, disk space and recovery

The existing GPT Realtime Assistant desktop workflow describes downloadable speech and general models on Apple silicon Mac. Treat each as its own requirement: a speech model turns audio into text, while a general model produces an answer. A translation workflow can depend on still another supported component.

This is not a claim that the public iPhone or iPad release offers the same local model system. Mobile online-session requirements remain separate. Check the installed Mac version, the controls actually offered and any stated hardware requirements before allocating storage or removing other files.

Use macOS System Settings > General > Storage to inspect available space and large categories. Keep room for normal work and temporary download or installation activity, not merely the final advertised model size. Exact assets and sizes vary, so use the size shown for the chosen model.

If the disk is nearly full, remove only files you recognize and can afford to lose or have verified elsewhere. Do not delete arbitrary Application Support folders, hidden caches or system data to make an installation fit. Apple’s storage guidance distinguishes user-managed categories from system-managed data.

Use the app’s model-management controls where provided, keep a stable connection and wait for a completed state. A partially downloaded file may occupy space without being loadable. Do not judge completion only by a file appearing in a folder.

Record the model name and readiness state. If a download fails, check available space and connection, then use the offered retry or removal control. Avoid launching several duplicate downloads. If the app does not explain its state, preserve the error and ask support rather than manually modifying a model directory.

After setup, close and reopen the app and select the local workflow. Test a new harmless spoken sentence, inspect its transcript and ask a new short question if local answers are supported. Then repeat with connectivity disabled. Previously saved text does not demonstrate fresh inference.

If transcription succeeds but answering fails, inspect the general model rather than redownloading speech assets automatically. If both fail before any output, check loading and compatibility. Keep a text alternative for important work until all the stages you need have passed the same test.

If the app provides a model removal control, use it for an unused model and confirm that disk space is reclaimed. Preserve the model you actually use until you know its replacement works. A smaller option may fit your device better, but compare task quality before relying on it.

Before travel, note which models need a future download to restore. Local assets do not automatically cover every language or every reasoning task, and account or subscription checks may have separate connectivity needs. A practical storage plan records what is installed, what was tested and what fallback remains.