Docs Start here

See it working

Videos and screenshots of BOAR on a real Android phone in airplane mode, with the default 1.5B model and answers that cite offline sources.

Everything here was recorded on a Xiaomi phone (MediaTek Dimensity, 12 GB RAM) in airplane mode, with the default model, Qwen2.5-1.5B-Instruct. Nothing left the phone.

The walkthrough

Four minutes, start to finish: airplane mode on, a first question with sources, a research prompt, a follow-up and a few facts. It was also posted on X.

Four clips

Offline, first question (0:35). Airplane mode on; a question answered with sources from the offline library.
A research prompt (1:17). A ready-made synthesis question from Prompt Ideas.
Follow-up (1:21). "Interesting, continue": the answer picks up from the conversation so far.
Facts and math (0:38). The capital of Australia from the Wikipedia pack, then 7×8.

Screens

An answer with its offline sources listed under it
Answers with sources. Every answer lists the offline articles it used and how well each matched. The OFFLINE badge and the airplane icon are real.
Prompt Ideas: ready-made research questions
Prompt Ideas. Ready-made research questions to try the app with.
The menu with past sessions, documents, settings and telemetry
Menu. Past sessions, your own documents, settings and telemetry. RAM and disk use are always visible at the bottom.
Installed models
Models. The default model and the embedding model, plus optional ones you can switch to with one tap.
More models and a Hugging Face search
Bring your own model. Mixture-of-experts and Gemma models tested on this phone, and a search for any GGUF model on Hugging Face.
Execution telemetry for each answer
Execution telemetry. Every answer is measured: tokens per second, time to first token, peak memory. Export as JSON or CSV.
Memory and storage use
Memory and storage. The app's RAM use against a 12 GB limit, and its storage against a 50 GB budget.
The last benchmark and the engine in use
Engine. The last answer's numbers, and the model, context size, threads and license in use (llama.cpp through llama.rn).

Measured on the phone

Six knowledge questions with the default model and the Wikipedia Vital Articles pack (2026-09-24): 6 of 6 answered correctly, all from retrieved articles, about 10.6 seconds per answer (6.4 s to the first word, 17.5 tokens a second), 1.76 GB peak memory. The full numbers are in Results, and you can run the same benchmark on your phone.

Edit this page on GitHub · Something wrong? Open an issue