RunAnywhere Flutter SDK

On-device AI for Flutter — run LLMs, speech-to-text, text-to-speech, and voice pipelines locally. Private, offline-capable, production-ready.

Requirements

Component	Minimum	Recommended
Flutter	3.44+	3.44.6 (see `pubspec.yaml`)
Dart	3.12+	3.12.2 (workspace constraint)
iOS	17.5+	17.5+
Android	API 24 (7.0+)	API 28+
Xcode	26+	26+
RAM	2 GB	4 GB+ for larger models

ARM64 devices are recommended. Metal on iOS and NEON on Android provide significant speedups.

Platform setup: iOS requires static CocoaPods linkage and microphone permissions for voice features. See docs/DEVELOPMENT.md for Podfile and manifest details.

Installation

Add the packages you need to pubspec.yaml. Pin to 0.20.12:

Core + LlamaCpp (LLM):

yaml

dependencies:
  runanywhere: 0.20.12
  runanywhere_llamacpp: 0.20.12

Core + ONNX (STT / TTS / VAD):

yaml

dependencies:
  runanywhere: 0.20.12
  runanywhere_onnx: 0.20.12

Core + MLX (Apple LLM / VLM / speech on physical iOS devices):

yaml

dependencies:
  runanywhere: 0.20.12
  runanywhere_mlx: 0.20.12

Optional NPU (Snapdragon Android):

yaml

  runanywhere_qhexrt: 0.20.12

Then run:

bash

flutter pub get
cd ios && pod install && cd ..   # iOS only

Package	Purpose
`runanywhere`	Core SDK — required
`runanywhere_llamacpp`	LLM / VLM (GGUF)
`runanywhere_onnx`	STT, TTS, VAD
`runanywhere_mlx`	Apple MLX
`runanywhere_qhexrt`	Qualcomm Hexagon NPU (optional)

Quick Start

The public API is namespaced — each capability is accessed via a static accessor on RunAnywhere:

dart

import 'package:runanywhere/runanywhere.dart';
import 'package:runanywhere_llamacpp/runanywhere_llamacpp.dart';

void main() async {
  WidgetsFlutterBinding.ensureInitialized();

  await RunAnywhere.initialize();
  LlamaCpp.register();

  await RunAnywhere.downloadModel('smollm2-360m');
  await RunAnywhere.llm.load('smollm2-360m');

  final response = await RunAnywhere.llm.chat('What is the capital of France?');
  print(response);
}

For streaming, voice, vision, RAG, and tool calling, see the documentation.

Production initialization

dart

await RunAnywhere.initialize(
  apiKey: '<YOUR_API_KEY>',
  baseURL: 'https://api.runanywhere.ai',
  environment: SDKEnvironment.SDK_ENVIRONMENT_PRODUCTION,
);

Capabilities

Accessor	Purpose	Example
`RunAnywhere.llm`	Text generation	`llm.chat()`, `llm.generate()`, `llm.generateStream()`
`RunAnywhere.stt`	Speech-to-text	`stt.transcribe()`, `stt.transcribeStream()`
`RunAnywhere.tts`	Text-to-speech	`tts.synthesize()`, `tts.speak()`
`RunAnywhere.vad`	Voice activity detection	`vad.detectVoiceActivity()`, `vad.streamVAD()`
`RunAnywhere.vlm`	Vision-language	`vlm.processImage()`
`RunAnywhere.voice`	Voice agent pipeline	`voice.initializeWithLoadedModels()`
`RunAnywhere.models`	Model registry	`models.register()`, `models.available()`
`RunAnywhere.downloads`	Download lifecycle	`downloads.start()`, `downloads.delete()`
`RunAnywhere.rag`	Retrieval-augmented generation	`rag.query()`
`RunAnywhere.tools`	Tool / function calling	`tools.generateWithTools()`

Flat aliases such as RunAnywhere.downloadModel() and RunAnywhere.loadModel() mirror the Swift/Kotlin/RN APIs for cross-platform parity.

Supported model formats

Format	Use case	Backend
GGUF	LLM, some VLM	`runanywhere_llamacpp`
ONNX / Sherpa archives	STT, TTS, VAD	`runanywhere_onnx`
MLX bundles	LLM, VLM, speech on Apple silicon	`runanywhere_mlx`
QHexRT bundles	NPU models on Snapdragon	`runanywhere_qhexrt`

Built-in catalog models (e.g. smollm2-360m) are discovered automatically — no manual registration required for catalog entries.

Error Handling

Errors throw SDKException with a proto ErrorCode:

dart

try {
  final result = await RunAnywhere.llm.generate(
    'Hello!',
    LLMGenerationOptions(maxTokens: 64),
  );
} on SDKException catch (error) {
  print('SDK error [${error.errorCode}]: ${error.message}');
}

Documentation & Examples

Resource	Link
Docs site	docs.runanywhere.ai/flutter/introduction
API reference	docs/Documentation.md
Example app	examples/flutter/RunAnywhereAI/
Contributing / platform setup	docs/DEVELOPMENT.md

FAQ

Do I need an internet connection?
Only for initial model download. Inference runs entirely on-device afterward.

Is user data sent to the cloud?
No. All inference is local. Production mode may collect anonymous telemetry (configurable).

Can I use custom models?
Yes — register GGUF, ONNX/Sherpa, or MLX bundles via RunAnywhere.models.register().

How do I test Apple MLX?
MLX execution requires a physical arm64 iOS device. The simulator slice is for compile/link validation only.

Support

Documentation: docs.runanywhere.ai
Discord: discord.gg/N359FBbDVd
GitHub Issues: github.com/RunanywhereAI/runanywhere-sdks/issues
Email: [email protected]

License

RunAnywhere License. See LICENSE for details.