Running the model on your own device means your prompt is never transmitted, never retained by anyone, never subject to a policy change you did not read, and never sitting in a breach you did not cause. You pay for that with model size: a model that fits in a phone's memory is weaker at hard reasoning, long code and current facts than a model that lives in a data center. For most of what people actually type into a chatbot, that trade is worth taking.
What happens to a prompt in the cloud
When you type into a cloud assistant, the text leaves your machine before anything thinks about it. It travels to a server you do not control, is processed there, and is usually retained for some period under a policy that can change. That retention is not sinister on its own. It is simply an additional place your words exist, owned by someone else, reachable by their staff under their rules, reachable by a court under its rules, and reachable by an attacker if one gets in.
None of that is abstract for a lawyer drafting around a client matter, a clinician typing a case summary, a journalist checking a name, or anyone who has ever deleted a half-typed question rather than send it. The more interesting part is how often ordinary users do the same thing without calling it privacy. People self-censor with cloud assistants constantly, and they get worse answers because of it.
What you get back
The first thing is that the question stops being a transmission. There is no copy of it anywhere but your own disk, so there is nothing to retain, subpoena or leak. The second is availability: a model on your device does not care about signal, roaming, a captive portal, a rate limit or an outage, so it answers in a tunnel and on a plane exactly as it answers at your desk. The third is that the meter stops running. Inference on your own silicon has no marginal cost, which is why Inborn can be a one-time purchase instead of a subscription with a message quota.
What it costs you
This is where most local-AI marketing goes quiet, so here are the numbers. A 2B parameter model quantized to four bits occupies 1.28 GB and runs at roughly 5 to 6 tokens a second on a 2018 Android flagship, and considerably faster on anything current. A frontier cloud model is two to three orders of magnitude larger. That gap shows up exactly where you would expect: multi-step reasoning, long code, mathematics, obscure facts, and languages with little training data. It shows up much less on the things people do most, which are rewriting an email, summarizing a page, explaining a term, translating a paragraph, drafting an outline, and turning notes into prose.
So the honest rule is not that local is better. It is that local is better for anything private, anything offline, and anything routine, while the cloud is better for the hard tail. Knowing which one you are doing is the whole skill.
How Inborn does it
Inborn ships a 0.8B model inside the app, 533 MB, so the first chat works before any download finishes. It profiles your device and tells you the speed to expect from each larger model before you install it rather than after. When a question is outside what the running model is good at, it says so instead of bluffing, and offers the model that would do better.
The verification side matters as much as the engineering. On Android the release build does not declare the INTERNET permission, so the operating system will not let it open a socket. On iPhone, the App Privacy Report shows an empty row. You do not have to believe the paragraph above: you can check it from outside the app in under a minute, and the next post is the list of how.
NextHow to choose an on-device AI model · Verify it yourself