← MPS Learning Center

[ reference · FAQ ]

On-device, cloud AI, and MPS: frequently asked questions

Is on-device AI private by default?

No. Local inference can avoid sending a specific input to a remote model, but privacy also depends on the app’s collection, permissions, storage, telemetry, sync, backups, model updates, and access controls.

Is cloud AI always more capable?

Not for every task. A remote service may offer larger models, shared retrieval, or centralized operations, while a compact local model can be adequate or preferable for a constrained task. Define and test the task-specific quality threshold.

Does on-device AI work without internet?

Some local features can continue without a network, but the product may still require connectivity for installation, authentication, model download, updates, sync, or other features. State the actual offline boundary.

How should cost be compared?

Compare lifecycle cost: usage and service fees; hardware and device replacement; integration; model distribution; support; monitoring; security; network; and the cost of failure or exit.

What devices are required for on-device AI?

Requirements vary by model, runtime, task, and desired experience. Test memory, storage, OS support, acceleration, heat, power, accessibility, and performance on the least capable supported device.

Does MPS certify an AI system?

No. MPS is a claim-level provenance framework for nonfiction. Its Preflight and Auditor tools support review workflow; they are not security certification, legal advice, or primary-source verification.

[ MPS implementation library ]

Choose a boundary, then test it.

Learning Center