On-device AI vs cloud AI: what actually changes for you
Most explanations of this choice are about servers. The useful version is about you: what happens to your words, how fast the answer arrives, what it costs, and how much the assistant can do.
The two designs, one sentence each
Cloud AI: your message travels to a data center, a large model answers there, and the answer travels back.
On-device AI: your message stays on your phone, and a smaller model answers there.
What changes
| On-device | Cloud | |
|---|---|---|
| Where your words go | Nowhere. The model is on the phone. | To the provider's servers. |
| Account | Not needed for the model itself. | Usually required, to bill and to rate-limit. |
| Works offline | Yes. | No. |
| Latency | The answer starts at once and arrives at reading speed. | Depends on the network and on server load. |
| Model size | Fits in 8 GB of memory. | Hundreds of billions of parameters. |
| Best at | Personal logistics, drafting, summarising, reading images, acting on the phone. | Long research, expert domains, long-form writing. |
| Cost to run | Your battery. | The provider's electricity, billed to someone. |
| When the policy changes | Nothing about your data changes. | Whatever the new policy says. |
Privacy is the biggest difference and the least visible
When words leave your phone, their protection becomes a matter of contracts, employees, and jurisdictions. When they do not leave, none of that applies. In day-to-day use the two designs feel similar, which is exactly why the distinction matters: you cannot feel the difference, so you have to check it.
Capability is the honest cost
A model that fits in a phone is smaller. For most requests to an assistant the gap is small; for demanding research it is not. Anyone who tells you there is no tradeoff is selling something. The practical question is what you actually ask an assistant to do.
The middle path
Run locally by default, and reach for a cloud model on the occasions that need one, with your own key and an explicit consent step. This is how Orvena is built. The assistant and every phone tool are local and free. Cloud models are optional, use your own API key, and are gated per provider, with a second question before sensitive history would leave the phone.
A way to decide
Write down the last ten things you asked an assistant. If most were about your own life, your calendar, your photos, a message you were writing, a place nearby, then on-device is the right default: the data is personal and the tasks are well within a phone-sized model's reach. If most were deep research, you will want cloud access some of the time. Many people want both, which is why the middle path exists.
Orvena runs its model on the iPhone and lets you add a cloud model with your own key when a request calls for it.
Download on the App Store