Ask on a plane.
Get sources.
Search engines and chatbots stop when the signal does. This one keeps going, and shows you where every claim came from.
- No network permissionThe app has no way to reach the internet. A script checks every build.
- Two modelsA 1.7B planner writes the searches. A 4B or 35B answerer writes the reply.
- Every claim citedTap a [1] to read the exact passage it came from.
Built for no signal.
Flights, hikes, remote towns. The places you most need answers are the places the internet doesn’t reach.
Cited answers
Claims carry [1]-style links that open the source passage.
Offline by design
No INTERNET permission at all, checked on every build.
Exact arithmetic
Fuel, temperature, battery, distance and weight are converted by code, not guessed.
Honest about gaps
Anything without a source is labelled unsourced, with a warning beside it.
Five steps, all on the phone.
Nothing leaves the device. The models and the library are copied over USB, never downloaded by the app.
Question
You ask in plain language.
Planner
Qwen3‑1.7B turns it into up to three searches.
Retriever
SQLite full‑text search picks the best passages.
Answerer
Qwen3‑4B, or a 35B model streamed from storage.
Answer
Streamed text with citations and a Sources panel.
Qwen3‑4B
Fits mid‑range phones. 28.9 GB of storage in total.
Qwen3.6‑35B‑A3B
A model larger than the phone’s memory, streamed from flash by BigMoeOnEdge. 38.7 GB in total.
24.6 GB in your pocket.
Seven files, copied to the phone once. After that, no connection is ever needed.
The numbers, as logged.
Every figure comes from a log in the repository. Nothing here is estimated.
Honest caveats. The 54% is above the 50% bar by less than the graders’ margin of error. No 12 GB phone has been measured yet. A second set of 40 never‑tuned questions hasn’t been run. Known faults: answers cut off at the length limit, restaurant answers that decline to name places, and two location questions refused.
Where it goes next.
Each milestone ships with a logged measurement, the same rule as everything above.
Working prototype
- Search, citations, unit conversions and “near me”
- Two model sizes chosen by RAM, and a one‑command quickstart
- Signed v0.8.9 APK with an offline check on every build
Trust the quality
- Fix cut‑off answers and the restaurant and location refusals
- Run the held‑out 40‑question set and publish the scores
- Measure on a real 12 GB phone
Make it fast
- Make the fast model the default where quality holds
- Prefix caching and a tighter context budget
- Target: first word in under 30 s on mid‑range phones
Easy install
- F‑Droid and signed direct downloads
- Setup on the phone, without a computer
- A storage manager for library packs
Community packs
- An open pack format: medicine, repair, survival, regional guides
- Licence review and checksums for every pack
- Languages beyond English
$OFFR
A 35B AI that runs on your phone with zero internet. Cited answers, no cloud.
0xee9069e8f520652e59674fad2120f6dea2671ba3
Chart on DexScreener
Cited answers
Every claim links to the source passage it came from.
Zero internet
The app declares no network permission. It works in airplane mode.
No cloud
Models and library live on your phone. Nothing is sent anywhere.
Always check the contract address above against the chart link before trading. The app works fully without the token. This is not investment advice, and nothing here promises returns or price movement. Tokens can lose all their value.
Running in minutes.
You need git, curl, bash and adb on a computer, an arm64 phone with USB debugging on, and 32–39 GB free on the phone.
git clone https://github.com/SharkCube99/offline-research-android.git
cd offline-research-android
scripts/quickstart.sh
# then switch on airplane mode and ask a question