Maathai llama.cpp Flutter Plugin #

This plugin packages llama.cpp so Flutter apps can load GGUF models and chat fully offline. The initial release focuses on Android (NDK-based build), with room to extend to other targets.

Getting Started #

Install the package:

flutter pub add maathai_llamma

Then initialize the plugin before loading a model:

final llama = MaathaiLlamma();
final ok = await llama.initialize();
if (!ok) {
  throw Exception('Failed to initialize llama backend');
}
await llama.loadModel(modelPath: '/sdcard/Download/mistral.q4_k.gguf');
final output = await llama.generate(prompt: 'Hello!');
print(output);

To stream tokens, call generateStream instead of generate to receive incremental chunks.

Directory Layout #

• lib/: Dart API exposing initialize → loadModel → generate → release lifecycle.
• android/: Native bridge (maathai_llamma_bridge.cpp) builds llama.cpp and exposes JNI entry points.
• extern/llama.cpp: Git submodule tracking upstream inference runtime.
• example/: Flutter UI that lets you pick a local model path and exchange prompts.

Platform Support #

Android (stable) #

1. Android Studio / SDK with NDK 26.1.10909125 and CMake 3.22.1.
2. Flutter 3.3+ with compatible Dart SDK.
3. GGUF model stored locally on the device/emulator.

iOS #

iOS support is currently a stub implementation that responds with an unimplemented error. Native bindings will ship in a future update.

Build Steps #

git clone https://github.com/your-org/maathai_llamma.git
cd maathai_llamma
git submodule update --init --recursive
flutter pub get

To validate the plugin and example:

cd example
flutter run --device-id <your-device>

The build will compile llama.cpp for each Android ABI via CMake and bundle the shared library libmaathai_llamma.so.

Runtime Workflow #

1. MaathaiLlamma.initialize() — calls llama_backend_init once per process.
2. loadModel(modelPath, contextLength, threads, gpuLayers) — loads a GGUF model, configures context size & thread count, and prepares a sampler chain (temperature + top-k/top-p).
3. generate(prompt, {maxTokens}) — sends the prompt to the native side, runs a decoding loop, and returns sampled tokens as a string. The Dart API now defaults maxTokens to 512, but passing 0 (or any value ≤0) lets the model continue until EOS or until an internal safety cap (1024 tokens or the remaining context, whichever is smaller).
4. release() — frees model, context, and sampler.

See example/lib/main.dart for an end-to-end chat UI.

Android Integration Tips #

• The plugin builds against c++_shared. If your host app already includes another STL, align them to avoid duplicate symbols.
• Ensure ABI folders (android/src/main/jniLibs/<abi>) contain libc++_shared.so or rely on Gradle packaging from the NDK.
• For large models, expect high memory pressure. Prefer quantized (Q4/Q5) variants for mid-range devices.

Roadmap #

• ❌ iOS/macOS support via upstream XCFramework.
• ❌ Streaming token events via EventChannel.
• ❌ Model metadata inspection in Dart.
• ❌ Integration tests with on-device smoke model.

Contributions welcome! Open an issue or PR to discuss enhancements.

Licensing #

• Core code in this repository is provided under the Maathai Community License v1.0, which is based on Apache-2.0 but limited to Non-Commercial Use below a $1,000,000 USD annual revenue threshold.
• For Commercial Use at or above the threshold, please obtain a commercial license from Maathai. Contact: support@usemaathai.com or visit https://usemaathai.com.
• Third-party components (e.g., extern/llama.cpp) are licensed under their respective terms and are not affected by the non-commercial clause. See THIRD_PARTY_NOTICES.md.