LlamaDroid

On-device LLM inference for Android — local, private, full control.

Run large language models entirely on your phone. No internet required after model download. No accounts, no telemetry, no data leaves your device.

Features

• Chat — streaming token-by-token generation, markdown rendering, conversation management
• Model Hub — browse and download GGUF models directly from HuggingFace with progress notifications
• System Prompts — 6 built-in presets (Creative, Precise, Code, Translator, Summarizer, Default) + custom
• Full Parameter Suite — temperature, top-p/k, mirostat v2, min-p, repeat/frequency/presence penalties
• Performance Profiles — one-tap presets (Creative / Precise / Fast)
• RAG — import TXT/MD documents, auto-chunk, keyword search, context-injected chat with citations
• Server Mode — OpenAI-compatible API with live metrics, Quick Settings tile, public/LAN access
• Multi-modal — image attachment for vision models, speech-to-text, text-to-speech
• Speculative Decoding — draft model support with configurable lookahead
• LoRA Adapters — import and load/unload adapters at runtime
• Benchmark — measure prompt/generation speed, memory usage
• Conversations — save, browse, export/import as JSON
• Onboarding — first-launch wizard with device tier detection and model recommendations

Quick Start

git clone --recurse-submodules https://github.com/projectamazonph/llamadroid
cd llamadroid
git submodule update --init --depth 1   # fetch llama.cpp
./gradlew assembleDebug

Prerequisites

• Android Studio Koala+ (2024.1+)
• NDK 27+ (via SDK Manager)
• CMake 3.22+
• Java 17+
• Android SDK 35+

Architecture

┌─────────────────────────────────────┐
│  Jetpack Compose UI (Material 3)    │
├─────────────────────────────────────┤
│  MVVM + Hilt DI + Room DB           │
├─────────────────────────────────────┤
│  InferenceEngine (Kotlin)           │
├─────────────────────────────────────┤
│  JNI Bridge (C++20)                  │
├─────────────────────────────────────┤
│  llama.cpp (Vulkan/OpenCL/NEON)      │
└─────────────────────────────────────┘

61 Kotlin files across clean architecture layers (data/domain/ui), 6 C++ bridge files with full inference, sampling, embeddings, and model info implementation.

Project Layout

app/src/main/
├── java/com/llamadroid/
│   ├── core/           DI modules, preferences
│   ├── data/           Room entities/DAOs, HF Hub API, JNI bridge
│   ├── domain/         Chat, inference, server, RAG, media engines
│   ├── service/        Foreground services + Quick Settings tile
│   └── ui/             Compose screens, components, navigation, theme
├── cpp/                JNI bridge + llama.cpp submodule
└── res/                Resources, manifest, file paths

Build Variants

Variant	ABIs	Use Case
debug	arm64-v8a, x86_64	Development
release	arm64-v8a, armeabi-v7a, x86_64	Production

License

MIT

Tech Stack

Layer	Technology
Language	Kotlin + Java
UI	Jetpack Compose (Material 3) + MVI
DI	Hilt
Async	Kotlin Coroutines + Flow
NDK	JNI Bridge (C++20)
LLM	llama.cpp (Vulkan / OpenCL / NEON)
Models	GGUF via HuggingFace
Build	Gradle (Kotlin DSL)
Build Variants	debug (arm64-v8a, x86_64) - release (arm64-v8a, armeabi-v7a, x86_64)

Quick Install

git clone --recurse-submodules https://github.com/projectamazonph/llamadroid
cd llamadroid
git submodule update --init --depth 1  # fetch llama.cpp
./gradlew assembleDebug

📊 Codegraph

See codegraphs/llamadroid.md for the full dependency graph.