Android-Assistant (llama_cli)

Minimal Android NDK build of a llama.cpp-based CLI that runs an interactive chat loop on-device. This repo builds a single native binary plus the NDK libc++_shared.so runtime.

Requirements

• Android NDK installed (r21+ should work; uses LLVM toolchain)
• adb for pushing/running on a device or emulator
• A GGUF model file compatible with llama.cpp

Build

1. Set the NDK path (one of these is required). Example:

export ANDROID_NDK_HOME=/path/to/android-ndk

You can also use ANDROID_NDK_ROOT or NDK_HOME instead of ANDROID_NDK_HOME.

2. Optional build tuning (defaults shown):

export ANDROID_ABI=arm64-v8a
export ANDROID_API=21

3. Build:

make

Outputs:

• build/llama_cli
• build/libc++_shared.so

Run (device/emulator)

1. Push the binaries and model:

adb push build/llama_cli /data/local/tmp/
adb push build/libc++_shared.so /data/local/tmp/
adb push /path/to/model.gguf /data/local/tmp/model.gguf

2. Run on device:

adb shell
cd /data/local/tmp
chmod 755 llama_cli
export LD_LIBRARY_PATH=/data/local/tmp
./llama_cli

3. Type a prompt and press Enter. Type exit to quit.

Configuration Model path is currently fixed at compile time via include/config.h. The default is MODEL_PATH "/data/local/tmp/model.gguf".

You can also tune N_CTX, N_BATCH, N_THREADS, and N_GPU_LAYERS in include/config.h.

If you want -m /path/to/model.gguf to work at runtime, update native/llama_cli.cpp to pass the CLI argument into load_model(...) instead of MODEL_PATH.