The Android app is the web UI in a WebView over an embedded
llmman serve: the same daemon, cross-compiled for aarch64-linux-android
and shipped with llama.cpp's own Android build, running on the phone. Every
model runs on-device; nothing in the app talks to a server that is not
127.0.0.1:17434. Source is android/.
Requirements: Android 9 (API 28) or newer, 64-bit ARM — the one Android build llama.cpp publishes. Inference is CPU-only, so pick models by RAM: a phone with 8 GB runs 1–4B parameter models at Q4 comfortably; 7–8B at Q4 needs 12 GB or more.
Download llmman-aarch64-linux-android.apk from the
latest release and
open it (sideloading has to be allowed for the browser or file manager
that opens it). The checksums.txt in the same release covers it.
The app is not on Google Play. Releases are signed with one key, so a newer APK installs over an older one.
On launch the app starts a foreground service that runs llmman serve and
shows the web UI once /api/version answers. The service keeps the daemon
up — a model stays loaded, a pull keeps going — with the app swiped away;
the notification's Stop ends it. If the daemon exits, the service restarts
it with backoff and the UI reloads. A failure shows the daemon's log in the
app.
Everything the web UI can do on a desktop works here: the Models page's
search, and pulling any reference llmman pull takes (docker.io/ai/…,
hf.co/…, …), chat
with streaming replies, the model picker, Export chats (into Downloads;
on Android 9 a file picker asks where), and Shell — a /system/bin/sh in the app's sandbox, with
llmman on PATH. Diffusion models are not supported: the mediagen
backend needs GPU libraries the phone lacks.
The daemon binds loopback only, which keeps the network out but not other
apps on the phone — Android's loopback is shared. So it runs with an API
key (api.md) generated once per install into the
app's private files: the WebView is handed it before the page loads, the
Shell's llmman has it as LLMMAN_API_KEY, and any other app that
connects to 127.0.0.1:17434 gets a 401. The Shell's token
(webui.md) changes with each daemon
start: the app reads it from the daemon's data directory and opens the UI
with it in the address's fragment. There is no setting to expose
the API or change the bind address.
Model storage is the app's private data (Android/data is not used) and
is removed with the app or by Clear storage. Conversations live in the
WebView's IndexedDB, as in a browser.
Needs, in addition to the usual build tools (Rust,
Go 1.27.3+): a JDK 17, the Android SDK with platforms;android-35,
build-tools;35.0.0 and ndk;27.2.12479018 (Android Studio or
sdkmanager installs them), rustup target add aarch64-linux-android and
cargo install cargo-ndk.
cd android
echo "sdk.dir=$ANDROID_HOME" > local.properties # or set ANDROID_HOME
./gradlew assembleDebug # app/build/outputs/apk/debug/app-debug.apk
./gradlew assembleRelease # debug-signed unless LLMMAN_ANDROID_* are setGradle does all of it: cargo ndk -t arm64-v8a --platform 28 build --release at the repository root, a download of the pinned
(LLAMA_CPP_RELEASE) llama-<tag>-bin-android-arm64.tar.gz, and staging
of both into the APK's native libraries. The version is Cargo.toml's; in
CI packaging/version.sh --apply sets it first, as for every other
release asset.
To sign a release with your own key, set LLMMAN_ANDROID_KEYSTORE (the
keystore, base64), LLMMAN_ANDROID_KEYSTORE_PASSWORD,
LLMMAN_ANDROID_KEY_ALIAS and LLMMAN_ANDROID_KEY_PASSWORD in the
environment; without them assembleRelease signs with the debug key.
Android forbids an app from executing anything under its writable storage
(since API 29), so every executable has to arrive through the APK's
native-library directory, which only takes lib*.so. The daemon is
therefore libllmman.so, llama.cpp's llama-server launcher is
libllama-server.so, and extractNativeLibs is on so they exist as files.
The service exposes both on PATH through symlinks named llama-server
and llmman in the app's files directory (the kernel resolves the link,
SELinux checks the target) and runs the daemon with LLMMAN_RUNTIME=path,
LD_LIBRARY_PATH set to that directory (llama.cpp's Android objects carry
no RUNPATH), HOME under the app's files directory and
LLMMAN_SHELL=/system/bin/sh.
Go has no c-archive build mode for Android, so there the Go shim is a
shared object (libllmman_shim.so, -buildmode=c-shared) next to the
binary, found through a $ORIGIN RUNPATH — build.rs switches on
CARGO_CFG_TARGET_OS. Every object is 16 KiB page-aligned, as Android 15
devices with 16 KiB pages require.
The same aarch64-linux-android binary runs outside the app too, e.g. in
Termux: --runtime bin downloads llama.cpp's Android release there, since
src/llama_release.rs knows its asset name.