Start with where the model processes each request
On-device inference runs a model on the phone or computer. Cloud inference sends a request to remote computing infrastructure. Some products route different tasks to different places, so the right question is not simply whether an app is 'local' or 'cloud-based'; ask where a specific conversation, image, or memory request is processed.
Apple's developer documentation gives one dated product example: it describes an on-device Apple Foundation Model and a Private Cloud Compute option with different context and reasoning capabilities. This illustrates a hybrid architecture, but Apple's privacy and security commitments apply to its own system and should not be generalized to unrelated companion apps.
Offline use, context size, and capability can pull in different directions
A local model can remain available without a network and may keep a request on the device, subject to the app's implementation. Its response quality and context budget depend on the model and hardware. A cloud model can use remote compute for larger or more complex tasks, but requires connectivity and a clear account of how requests are handled.
In Apple's current comparison, the on-device model works offline and has a smaller context window; the Private Cloud Compute model requires a network and offers a larger context window. Those are Apple-specific figures, not a universal rule that every cloud model is stronger or every local model is more private.
A privacy promise needs a concrete architecture behind it
Apple describes Private Cloud Compute as processing requests on specific server software that can be inspected and that is designed not to retain request data after the response. That is a provider's description of a particular architecture, with technical documentation available for scrutiny. A generic 'private AI' label from another app does not inherit the same design or assurances.
Also separate model inference from the rest of the service. Chat logs, account identifiers, payment events, crash reports, and backups may follow different data paths. Read the companion's privacy notice and ask what happens when you delete a conversation or account, even if its model can run locally for some tasks.
A practical checklist for choosing a companion's processing model
Match the architecture to your routine. If you need offline access, verify that the particular feature works offline. If your roleplay relies on long context, check the actual context and memory behavior. If data location matters, ask which requests leave the device, what is retained, and whether the answer applies to chats, images, and stored memory alike.
No architecture eliminates every trade-off. Look for settings that disclose processing, controls to delete or manage saved information, and a way to use the companion comfortably without sharing details you would rather keep private. The Jess site describes its current Discord-based experience; its published information should be reviewed separately from Apple or any other model provider.
Sources and further reading
- Apple Developer — adding server-side intelligence with PCC
- Apple Security — Private Cloud Compute Security Guide
Research and product details were reviewed September 30, 2026 where applicable. Vendor policies, plans, and features can change; verify current details with the linked first-party source.