Run 1 · Wikipedia Vital Articles pack
One average answer, second by second
0 s6.4 s10.6 s
- Answered correctly
- 6/6
- First token
- 6.4 s avg
- Generation
- 17.5 tok/s avg
- Whole answer
- 10.6 s avg
- Peak memory
- 1.76 GB
- Network
- none
How it works
Download once. After that, it’s yours.
BOAR needs the internet one time, to fetch its model. From then on, everything it knows and everything it says stays on your phone.
-
01
Install and download once
Get
boar-v1.0.0-arm64.apkfrom the release and check its SHA-256. On first launch BOAR fetches its default model and a small embedding model: about 1 GB, the only time it needs the network. -
02
Go where the signal doesn’t
Airplane mode on, Wi-Fi off, mobile data off. BOAR searches its offline library with keywords and on-device embeddings, and runs the model locally.
-
03
Ask, and see the sources
Every answer lists the offline articles it was built from, so you can check it. And every answer is measured: model, time to first token, speed, memory.
- 49,832 articles, 164 MBThe optional Wikipedia Vital Articles pack, on top of the built-in library.
- Your own knowledgeImport .txt, .md, .csv, .json and .pdf files as collections you can switch on and off, and share them as portable packs.
- Your own modelSearch Hugging Face for another GGUF model and run it on the phone.
- No account, no serverNo sign-in, no API and no Play Services or Firebase. No server ever hears your questions, so there is nothing to log or switch off.
When it matters
The moments you can’t look things up.
The six places in the film. BOAR reasons with a model on the phone and grounds what it says in its library. Each source named here is an article in the Wikipedia Vital Articles pack BOAR can download. Answers are abridged and written for the page. The train question is from BOAR’s evaluation set, where the default model answered 6:15 pm on a real phone.
-
In flight
The man next to me is choking and can’t speak. What do I do?
BOARLean him forward and give up to five firm blows between the shoulder blades. If that fails, give up to five abdominal thrusts. Keep alternating, and call the crew.
Choking
-
Underground
A train leaves at 3:40 pm and the trip takes 2 hours and 35 minutes. What time does it arrive?
BOAR3:40 pm plus 2 hours is 5:40 pm. Add 35 minutes: 20 to reach 6:00, then 15 more. 6:15 pm.
Worked out on the phone · no source needed
-
Mountain hut
My friend is shivering, clumsy and slurring words. Is it hypothermia?
BOARThose are classic signs. Get her out of the wind, swap wet clothes for dry layers, warm her core first, and give a warm sweet drink if she’s alert.
Hypothermia
-
In the garden
Why plant corn, beans and squash together?
BOARThey help each other. Corn gives the beans a pole to climb, beans add nitrogen to the soil, and squash leaves shade the ground and hold in moisture.
Companion planting
-
At sea
Thunder came 9 seconds after the flash. How far away is the storm?
BOARSound covers about 1 km every 3 seconds, so roughly 3 km. That’s close: head for shore or shelter now.
Thunder
Lightning
-
Flooded road
Water is running across the road ahead. Can I drive through it?
BOARDon’t. Moving water can float a car even when it looks shallow, and the road underneath may be gone. Turn around and wait on higher ground.
Flash flood
For anything serious, BOAR is a companion, not a doctor or a rescue service. Get help as soon as you can reach it.
The receipt
A model card isn't a benchmark.
Every answer records which model ran, how long until the first word, how fast it generated and how much memory it used. These are the first runs on a real phone, failures included.
Run 2 · five configurations, 17 questions each, same phone
Seconds per answer (median)
Show as a table
| Model | tok/s | First token | Per answer | What happened |
|---|---|---|---|---|
| Qwen2.5-1.5B (default) | 11.4 | 13.6 s | 21.0 s | 17/17, no timeouts |
| LFM2.5-8B-A1B (MoE) | 14.8 | 27.2 s | 57.3 s | Only model to solve the RAM question; 4 of 17 ran out of answer budget |
| Phi-3.5-mini | 4.0 | 44.0 s | 73.1 s | 4 timeouts at 120 s |
| Qwen2.5-7B | 2.7 | 70.5 s | 107.1 s | 5 timeouts at 120 s |
| Instella-MoE-16B-A3B | — | — | — | Didn't load: architecture not supported yet |
Medians. LFM2.5 ran separately with the same 17 questions. Highest peak memory of any run: 5.21 GB, under the 12 GB limit.
Measured on Xiaomi 2311DRK48G · MediaTek Dimensity 8300 · 11.6 GB RAM · Android 16. Raw results: docs/evidence.
Numbers look wrong? Run them on your phone and tell us. A result from your device is data nobody else has.
npm run eval:device -- --models <model>From the manifesto
Nothing leaves your phone.
Useful now beats perfect later.
No account, no API, no Google Play Services. One download at setup, then nothing. Not “mostly local”, not “offline except when it matters”.
Still growing
Not the dream, and we say so.
The model everyone wants on a phone, a huge mixture-of-experts model streaming from storage, doesn't exist yet. BOAR is what works today, measured honestly. 4–15 tok/s is enough to find out how to treat a blister, what a museum is about, or why the train isn't running, when there's no signal and nobody around to ask.
Open source
MIT licensed. Code, models, data sources and benchmark results are in the repo, and the APK ships with its SHA-256.
github.com/rferrari/boar-appStill open
- Not measured yet: a comparison against internet search plus frontier models.
- Not run yet: GrapheneOS on real hardware.
- Weak spots: synthesis on the 1.5B model, and citation formatting.
- English only, for now: the library and search are English; the menus are in English and Portuguese.
Where it started
Born from a poidh bounty. Vitalik asked for a research tool that lives on your phone and works with no signal, and poidh bounty #31 put it as a challenge: “Build the Best Offline AI Research App for Android”. BOAR is our answer, claim #124. The bounty is still open.
Bounties
Get paid to show it working.
Small tasks posted on poidh and paid in ETH: film BOAR answering with no signal, run a test on your phone, send the proof. Anyone with the app can claim one.
Roadmap
First the roots. Then what grows from them.
A direction, not a schedule. First, make the everyday app excellent, tested by the people who use it. Then BOAR grows into a library that the community fills and shares, one pack at a time.
-
Phase 1 · Shipped
Roots
Works today, first on Android: a model and a searchable library on your phone, your own documents, and any GGUF model you choose.
-
Phase 2 · Here today
First shoots
iOS is on the way. Packs for places and the right model for each task are in testing, and you can already ask out loud; spoken answers come next.
-
Phase 3 · After that
Branches
Packs people share: a community pack library, private encrypted memory, and essentials for first aid, disaster response and growing food, the first pack already in testing.
-
Phase 4 · Further out
Grove
Beyond the phone: a desktop app for Mac and Windows, and packs passed from phone to phone with no internet in between.
$boar · Base
Trade $boar right here.
Priced by SwapsPro, from whatever you're already holding — any EVM chain, Bitcoin, no bridge to find yourself. Connect only lives in this box; the rest of the page never touches your wallet.
$boar · Base 0x0cbf291Ba052174879d90bf781dF1A5F2BC5Bb07
Keep your phone charged, pack BOAR, and go somewhere without signal.
v1.0.0 · 123 MB · any 64-bit ARM Android phone · MIT · Read the manifesto