RunAnywhere AI for Flutter

A starter app for the RunAnywhere SDK, written in Dart.

The home screen is a grid of eight cards, one per feature. Each screen downloads and loads its own model the first time you open it, then runs it on the device. Copy a screen, point it at your own model, and you have the shape of a real app.

What each screen does

Card	SDK surface it calls	Model
Chat (text generation)	RunAnywhere.llm.generateStream, llm.cancel	Qwen3.5 0.8B
Vision (image understanding)	RunAnywhere.vlm.generateStream with ImageInput.file	SmolVLM 500M
Speech (speech to text)	RunAnywhere.stt.transcribe with AudioInput.pcm16	Whisper Tiny EN
Voice (text to speech)	RunAnywhere.tts.speak, tts.playbackState, tts.stop	Piper en_US lessac medium
Activity (voice detection)	RunAnywhere.vad.openStream plus the SDK's AudioCaptureManager	Silero VAD
Pipeline (voice agent)	RunAnywhere.voice.createSession(stt:, llm:, tts:)	Whisper Tiny + Qwen3.5 0.8B + Piper
Knowledge (RAG)	RunAnywhere.rag.open, session.ingest, session.query	all-MiniLM-L6-v2 + Qwen3.5 0.8B
Tools (function calling)	RunAnywhere.llm.tools.register, llm.generate with autoExecute: true	Qwen3.5 0.8B

Some behavior the table does not capture:

• Chat sends only the latest message. There is no conversation history or system prompt in the request, so the model is stateless across turns even though the UI looks like a thread.
• Speech is batch transcription. The app records with package:record into a PCM16 buffer and transcribes once on stop. There is no partial or streaming transcript.
• Pipeline hands the whole loop to the SDK. VoiceSession owns mic capture, turn segmentation, and TTS playback; the app only renders VoiceAgentStateChanged, VoiceUserTranscribed, and VoiceAgentResponse events.
• Knowledge ingests pasted text only. There is no file or PDF picker, and answers are shown without source citations.
• Tools registers three functions: get_weather, get_current_time, and calculate. The SDK runs the tool loop and the app supplies the executors.

Requirements

• Flutter 3.44.0 or newer, Dart 3.12.0 or newer
• iOS 17.5+ (the SDK podspecs set that deployment target and the Podfile pins it)
• Android API 24+; the SDK's Android modules compile against SDK 36 and NDK 28.2.13676358
• Xcode for iOS, Android Studio or the command-line SDK for Android

The Apple MLX backend needs a physical iOS device; MLX.register() returns false on the simulator. The Qualcomm Hexagon backend needs an arm64 Snapdragon device; the app checks QHexRT.isAvailable before registering it. Both are skipped cleanly elsewhere, so a plain simulator run works and falls back to llama.cpp and ONNX.

Running it

git clone https://github.com/RunanywhereAI/flutter-starter-example.git
cd flutter-starter-example
flutter pub get
flutter run

Microphone, camera, and photo library permissions are already declared in ios/Runner/Info.plist and android/app/src/main/AndroidManifest.xml. The Podfile already sets use_frameworks! :linkage => :static, which the SDK requires on iOS; without it the vendored xcframework symbols fail to link.

No API key or account is needed. RunAnywhere.initialize() is called with no arguments.

Models are not bundled. The first time you open a feature screen you get a download button, and the download runs over the network into local storage. Budget roughly 1.2 GB if you try every screen.

How the SDK is wired up

lib/main.dart does the setup in order: initialize the SDK, register the backends, register the model catalog, then start the app.

await RunAnywhere.initialize();

LlamaCpp.register();
await Onnx.register();
await MLX.register();
if (QHexRT.isAvailable) {
  await QHexRT.register();
}

await ModelService.registerDefaultModels();

ModelService (a ChangeNotifier behind provider) owns the catalog and the download and load state. It registers models with RunAnywhere.models.register using three registration shapes:

• ModelRegistration.url for single files (Qwen3.5, Silero VAD)
• ModelRegistration.archive for tarballs (SmolVLM, Whisper, Piper)
• ModelRegistration.multiFile for models that need companion files (LFM2.5-VL weights plus its mmproj projector, MiniLM model plus vocab)

Loaded-model state is mirrored into a synchronous Set<ModelCategory> because RunAnywhere.models.state() is async and widget build() needs an answer immediately. refreshLoadedModels() re-syncs it after every load or unload.

Project layout

lib/
  main.dart                     SDK init, backend registration, app root
  services/model_service.dart   Model catalog, download and load state
  theme/app_theme.dart          AppColors palette and AppTheme.darkTheme
  views/
    home_view.dart              Feature grid
    chat_view.dart              LLM streaming chat
    vision_view.dart            Image understanding (VLM)
    speech_to_text_view.dart    Recording and transcription
    text_to_speech_view.dart    Synthesis and playback
    vad_view.dart               Live speech detection
    voice_pipeline_view.dart    Voice agent session
    knowledge_view.dart         RAG over pasted text
    tool_calling_view.dart      Function calling
  widgets/
    feature_card.dart           Home grid tile
    model_loader_widget.dart    Download and load gate screen
    chat_message_bubble.dart    Message bubble with token metrics
    audio_visualizer.dart       Audio level bars

The app is dark theme only. AppTheme.darkTheme is Material 3 with Inter for body text and Space Grotesk for headings, both via google_fonts.

SDK packages

Pinned at ^0.20.19, which is the latest version of all five on pub.dev.

Package	Role
runanywhere	Core SDK: model management, inference APIs, FFI bridge to the C++ core
runanywhere_llamacpp	llama.cpp backend for LLM and VLM
runanywhere_onnx	Sherpa and ONNX backend for STT, TTS, VAD, and embeddings
runanywhere_mlx	Apple MLX backend, physical iOS devices only
runanywhere_qhexrt	Qualcomm Hexagon NPU backend, Android arm64 only

Models

Registered in ModelService.registerDefaultModels(). Sizes are the memoryRequirementBytes each model declares.

Every loader screen names the model it is about to fetch and who published it, so nothing downloads without saying what it is.

Model	Category	Size
Qwen3.5 0.8B Q4_K_M	Language	533 MB
SmolVLM 500M Instruct	Multimodal	600 MB
LFM2.5-VL 3B (Q4_K_M + mmproj)	Multimodal	2.26 GB
Sherpa Whisper Tiny EN	Speech recognition	75 MB
Piper TTS en_US lessac medium	Speech synthesis	65 MB
Silero VAD	Voice activity	2.3 MB
all-MiniLM-L6-v2	Embedding	25.5 MB

LFM2.5-VL is registered but no screen selects it. downloadAndLoadVLM takes a modelId that defaults to SmolVLM; pass ModelService.vlmLfm25ModelId to route the Vision screen through the larger model instead.

Adding your own model

Register it alongside the defaults in lib/services/model_service.dart. Registration is idempotent, so re-running it is safe.

await RunAnywhere.models.register(
  ModelRegistration.url(
    id: 'qwen2.5-1.5b-instruct-q4_k_m',
    name: 'Qwen2.5 1.5B Instruct',
    url: 'https://huggingface.co/.../qwen2.5-1.5b-instruct-q4_k_m.gguf',
    framework: InferenceFramework.INFERENCE_FRAMEWORK_LLAMA_CPP,
    category: ModelCategory.MODEL_CATEGORY_LANGUAGE,
    memoryRequirementBytes: 1100000000,
  ),
);

Then point a screen at the new id, either by changing the matching static const String at the top of ModelService or by passing the id through the download and load method.

Privacy

Inference runs on device. Prompts, audio, and images are not sent anywhere. Two things do use the network: model downloads, and the get_weather tool on the Tools screen, which calls the public Open-Meteo API with the location string the model extracted from your prompt.

The other apps

Platform	Repo
React Native	react-native-starter-app
iOS and macOS	runanywhere-ios
Android	runanywhere-android
Web	runanywhere-web
Windows	runanywhere-electron
SDK monorepo	runanywhere-sdks
Documentation	docs.runanywhere.ai
Discord	discord.gg/N359FBbDVd

License

The starter app is MIT licensed (see LICENSE). The RunAnywhere SDK ships under its own license, included with each package on pub.dev.

Bug reports: github.com/RunanywhereAI/runanywhere-sdks/issues