Private on-device AI · iPhone · Android · Web · Desktop
AI that stays
on your phone.
Airplane mode on. It still answers, same speed. Inborn is an AI chat app that runs on your own device: no account, no cloud, and nothing you type goes anywhere, which is something you can check in a minute rather than take on trust.
Opens Inborn in this browser with your message ready. Your browser downloads one model once: 533 MB (Instant) or 1.3 GB (Fast, on capable computers and tablets). It recommends the one your device runs best, you can pick the other, and then it answers with no network at all. The line you type travels in the web address to open the app, the way any link does. Nothing after it leaves your browser.
OUT 0 BCONNECTIONS 0SEALED · ON-DEVICE
What the app's own Proof screen shows on your device — a picture, not a live reading of this page.
An assistant that lives on the hardware you already own
Nothing leaves your device.
Your messages, your documents and the model itself sit on your device. There is no server to reach, no account to create, and no analytics, ads or crash reporting in the app. On Android it does not hold the INTERNET permission, so it cannot open a connection at all.
Works where nothing else does.
A plane, a basement, a train through a tunnel, a country where you would rather not be online. The model is already on the device, so it never waits on a network and never fails because of one.
Runs on your own chip.
Open-weight models in GGUF, sized to what your phone can carry. The app measures your device, tells you how fast each model runs on it, and says so when a bigger one would answer better.
Pay once, or not at all.
Free is the whole app, not a trial: unlimited chat, no message quota, no watermark. Pro is $19.99 one time. There is no renewal, and no month where it stops working because you stopped paying.
Three steps, and then no internet
-
Install it
Get it from your store like any app. There is no sign-up screen, no email, and no onboarding that asks who you are.
-
It brings its own model
On iPhone, the Instant model, 533 MB, is inside the app download. On Android, you choose a model on first launch and Google Play delivers it as an asset pack after one tap, the way it delivers an update.
-
Turn the internet off and ask
Switch on airplane mode and keep going. The answer comes at the same speed, because there was never a server in the path.
Your photos and PDFs, read on your phone
Read a photo. Offline.
Receipts, labels, screenshots. Read here, never sent. Attach a picture in a chat and the model looks at it on the device. Its photo pack downloads once, when you attach your first photo, and from then on a picture is read with the radios off. Photos are read in the iPhone and Android apps; the browser version cannot look at them. One photo per message on Free.
Ask your PDFs. Nothing uploaded.
Every answer cites the page. The file never leaves. Add a PDF, Word file, text or a scan: it is read, indexed and searched on the device, and the answer carries SOURCES chips like lease.pdf · p.3 that open the passage it came from. Free indexes one document of up to 20 pages per chat; Pro removes the limit and adds OCR for scans.
Lock it. Wipe it.
Screen lock, incognito chats, emergency wipe. The sections above are about the cloud not getting your data. This one is about the person holding your phone.
Face ID or your passcode opens Inborn.
Chats stay encrypted until then. On Android it is the fingerprint or face unlock the phone already has. While the lock is on, the app also hides your chats in the app switcher.
Not saved, no memory.
An incognito chat is gone when you close it: nothing written to the database, nothing remembered for the next chat, and a file you added inside it is removed with it.
Delete everything.
Chats, the encryption key and every document on the device, after one confirm step. Nothing was ever anywhere else, so nothing can be recovered. If you turn it on, too many wrong unlock attempts wipe as well.
Verify it yourself.
Live network meter, inside the app. OUT 0 B. Every AI app calls itself private; these are the checks that do not need our cooperation, three of them made from outside the app, on the build you installed.
The airplane test
Switch on airplane mode and ask it anything. It answers at the same speed, because the model is already on the device. A badge stays in the header for as long as the radios are off.
No internet permission
Open the Play listing, then About this app, then App permissions. "Have full network access" is not there. The release build never declares it, and our own build fails if it ever appears.
aapt2 dump permissions inborn.apkApp Privacy Report
Settings, Privacy & Security, App Privacy Report. iOS lists every domain each app contacted in the last 7 days. Inborn's row stays empty, unless you started a model download yourself.
The exit meter
The Proof screen reads the operating system's own counters: bytes out, bytes in, open connections this session. Under them, TRACKERS lists what is in the build: no analytics, crash, ads or attribution SDK. Bytes out stays at 0 B unless you start a model download yourself.
The firewall test, the complete list of when any byte can leave, and what these checks do not prove: the proof page.
A model that fits your phone.
Instant, Fast or Sharp. Speed measured on your phone. Bigger models write better and reason further; they also need more memory, drain more battery and answer more slowly. The vault measures your device, marks the models that fit it and prints the tok/s it measured under each one, so you see that trade on your exact phone before you install anything. Every model below is open-weight under a permissive license.
| Model | Size | License | Good at | On a 2018 Android flagship |
|---|---|---|---|---|
| Instant Qwen3.5 0.8B · Free | 0.53 GB | Apache-2.0 | Quick answers, rewrites, summaries. Weak at code and math. | 12-15 tok/s |
| Fast Qwen3.5 2B · Free | 1.28 GB | Apache-2.0 | The everyday one: writing, questions, translation into the major languages. | 5-6 tok/s |
| Sharp Qwen3.5 4B · Pro | 2.74 GB | Apache-2.0 | Analysis, code and long documents. Slower, and the phone runs warmer. | too slow to use |
| Sharp (Phi) Phi-4-mini 3.8B · Pro | 2.49 GB | MIT | Reasoning, math and code, in English. Weaker outside English. | 2-3 tok/s |
Speeds are ours, measured on our own devices, not estimates: a OnePlus 6T from 2018 for the column above, and about 36 tokens a second for Instant on an iPhone 13 Pro. A newer phone is faster. The app profiles your device and prints its own number under every answer. You can also import any GGUF file you already have, including Gemma, Mistral and Llama. A model this size is weaker than one on a rack of servers, which is the honest trade set out in the best local LLM for a phone, compared.
How much does Inborn cost?
Inborn is a one-time purchase, never a subscription. Free is a complete app: unlimited chat, no message limit, no watermark.
$0
A complete app, not a trial.
- Unlimited chat with the built-in model
- No message caps, no watermarks, no ads
- One document of up to 20 pages per chat, one photo per message
- Proof screen, incognito, screen lock, wipe
$19.99one time
Buy it once on your store and keep it. One month of a cloud AI subscription, and it never renews.
- Unlimited documents and OCR for scans, every answer citing the page
- Voice in and out, with voices that run locally
- The larger Sharp models, unlimited personas, folders
- Export everything to a file you control
$69.99one time
For people whose notes must not leave the room. $49.99 as an upgrade from Pro.
- Client vaults, kept apart from everything else
- An audit log and signed export
- An architecture statement to hand to your firm
List prices in US dollars; local prices vary. A purchase is tied to the store you bought it from, because there is no account that could link them. Details in the terms. Inborn is a one-time purchase, and here is how Inborn compares with ChatGPT and Ollama, priced and licensed side by side.
Questions people ask
Is Inborn really offline?
Yes. Inborn runs the AI model on your own device, so it answers with every radio switched off, and the Android release build does not declare the INTERNET permission, which means it cannot open a connection. Turn on airplane mode and ask it something: the answer arrives at the same speed, because there is no server in the path.
Do you see my chats?
No. Your messages never leave the device, and Inborn has no account system, no server, no analytics and no crash reporting, so there is nothing for us to see, store, sell or hand to anyone. Conversations live in an encrypted database whose key the operating system holds. We could not recover them for you even if you asked.
How big is the download?
The app is a normal store download. On iPhone, the Instant model, 533 MB, ships inside the app. On Android, you choose a model on first launch and Google Play delivers it as an asset pack after one tap. Larger models are optional: Fast is 1.28 GB and Sharp is 2.74 GB, and you choose when to install them. In a browser, one model downloads once into the browser's own storage: 533 MB (Instant) or 1.3 GB (Fast, on capable computers and tablets), whichever you pick, with the one your device runs best offered first.
Which AI models does Inborn run?
Inborn runs open-weight models in GGUF format: Qwen3.5 at 0.8B, 2B and 4B under Apache-2.0, and Microsoft's Phi-4-mini 3.8B under MIT. Instant and Fast are free; the Sharp models come with Pro. You can also import any GGUF file you already have, including Gemma, Mistral and Llama.
Will it work on my phone?
Inborn measures your device and recommends a model that fits it, and the built-in Instant model runs on every supported device. Measured on our own hardware: Instant reaches about 36 tokens a second on an iPhone 13 Pro and 12 to 15 on a 2018 Android flagship. On that 2018 phone the Sharp models are too slow to be worth installing, and the app says so on the card before you install.
Is Inborn a subscription?
No. Inborn is a one-time purchase: Free is the complete app with unlimited chat, Pro is 19.99 USD once and Work is 69.99 USD once, with no renewal, no monthly fee and no message quota. List prices are in US dollars and local prices vary. A purchase belongs to the store account that bought it.
How good is a small on-device model compared with ChatGPT?
A model that fits on a phone is genuinely weaker than a frontier cloud model at hard reasoning, long code, current facts and rare languages, and it is the better choice for anything you would rather not upload. Expect solid short answers, rewrites, summaries and translation into the major languages. Expect mistakes elsewhere, and check anything that matters.
Can Inborn read my own documents?
Yes. Inborn indexes your PDFs and documents on the device and answers with the page it took the answer from, as SOURCES chips under the answer, and nothing is uploaded because there is no cloud index. Free reads one document of up to 20 pages per chat; Pro removes the limit and adds OCR for scans. A file you add inside an incognito chat is removed when the session ends.
Can I use Inborn for client work?
Yes. Work is a one-time 69.99 USD purchase that adds client vaults, an audit log with signed export, and an architecture statement you can hand to whoever at your firm has to approve software. Because there is no server, client material never leaves the machine you typed it on. You stay responsible for checking AI output under your own professional rules.
Is Inborn open source?
No, it's source-available. The code is public at github.com/moshecohen90/inborn, so you can read it, build it and publish what you find, but the license does not let you redistribute or modify it, which is what open source would require. None of the checks we publish depends on that: every one of them is made from outside the app, on the build you installed, and the Proof screen names the exact commit that build came from.
Get Inborn
iPhone and Android first. Windows and macOS after them. The browser version runs from the box at the top of this page.
Neither store listing is public yet; both links open the real product page the moment it is. Inborn's source is public at github.com/moshecohen90/inborn, source-available under the existing license: read it, build it, publish what you find. There is still no public issue tracker: support is an email to a person.