RunAnywhere Swift SDK

Run large language models, speech recognition, text-to-speech, and voice agents directly on iPhone, iPad, and Mac — with low latency and on-device privacy. The Swift SDK provides a unified API over pluggable backends (llama.cpp, ONNX, Apple MLX) so you can ship AI features without sending user data to the cloud for inference.

Requirements

Requirement	Minimum
iOS	17.5+
macOS	14.5+
Swift	6.2
Xcode	26+

Installation

Swift Package Manager

In Xcode: File → Add Package Dependencies… and enter:

https://github.com/RunanywhereAI/runanywhere-sdks

Select version 0.20.11 (or from: "0.20.11"), then add the products you need. Resolve against the newest tag on the Releases page if 0.20.11 binary assets are not published yet.

Product	Required	Capabilities
`RunAnywhere`	Yes	Core SDK, model lifecycle, events
`RunAnywhereLlamaCPP`	For LLM/VLM	GGUF models via llama.cpp + Metal
`RunAnywhereONNX`	For STT/TTS/VAD	Whisper, Piper, Sherpa-ONNX
`RunAnywhereMLX`	For MLX models	Apple MLX LLM, VLM, STT, TTS

Package.swift

swift

dependencies: [
    .package(url: "https://github.com/RunanywhereAI/runanywhere-sdks", from: "0.20.11")
],
targets: [
    .target(
        name: "YourApp",
        dependencies: [
            .product(name: "RunAnywhere", package: "runanywhere-sdks"),
            .product(name: "RunAnywhereLlamaCPP", package: "runanywhere-sdks"),
            .product(name: "RunAnywhereONNX", package: "runanywhere-sdks"),
        ]
    )
]

Quick Start

swift

import RunAnywhere
import LlamaCPPRuntime

@main
struct MyApp: App {
    init() {
        Task { @MainActor in
            // 1. Register the LLM backend
            LlamaCPP.register()

            // 2. Initialize the SDK
            try RunAnywhere.initialize(
                apiKey: "<YOUR_API_KEY>",
                baseUrl: "https://api.runanywhere.ai",
                environment: .production
            )

            // 3. Generate text. The model auto-loads (and downloads when
            //    absent); call `RunAnywhere.models.load(id:)` first if you
            //    want to control when that cost is paid.
            let result = try await RunAnywhere.llm.generate(
                prompt: "What is the capital of France?",
                options: LlmOptions(model: "llama-3.2-1b-instruct-q4", maxOutputTokens: 128)
            )
            print(result.text)
        }
    }

    var body: some Scene {
        WindowGroup { ContentView() }
    }
}

For development without an API key, call try RunAnywhere.initialize(environment: .development).

Capabilities

Capability	Backends	Model formats	Notes
LLM	`RunAnywhereLlamaCPP`, `RunAnywhereMLX`	GGUF (`.gguf`)	Streaming, structured output, tool calling
STT	`RunAnywhereONNX`, `RunAnywhereMLX`	ONNX (`.onnx`, `.ort`)	Whisper, Sherpa streaming
TTS	`RunAnywhereONNX`, `RunAnywhereMLX`	ONNX, Piper voices	System TTS also available
VLM	`RunAnywhereLlamaCPP`, `RunAnywhereMLX`	GGUF multimodal	Image + text inference
Voice	Core + STT + LLM + TTS backends	—	VAD → STT → LLM → TTS pipeline
NPU	—	—	Not available on Apple platforms via this SDK

swift

LlamaCPP.register()  // LLM + VLM
ONNX.register()      // STT + TTS + VAD (requires RunAnywhereONNX)
MLX.register()       // MLX modalities (requires RunAnywhereMLX)

Supported model formats

Format	Extension	Backend	Use case
GGUF	`.gguf`	llama.cpp	LLM, VLM text generation
ONNX	`.onnx`	ONNX Runtime	STT, TTS, VAD
ORT	`.ort`	ONNX Runtime	Optimized STT/TTS

Models are discovered through the RunAnywhere catalog, downloaded on-device, and managed via RunAnywhere.models.load(id:), RunAnywhere.models.download(id:), and the rest of the models namespace.

Connect (trusted LAN)

ConnectSession publishes or joins a language-model session on the local network. Commons owns role policy and protocol validation; this SDK owns Bonjour and the framed TCP adapter.

Role	Platforms
Host	macOS only
Client	iOS and iPadOS (Android clients use the Kotlin SDK)

Service type: _runanywhere-connect._tcp
Hosting is app-scoped: the macOS app must keep the model loaded and the process alive
Clients adopt the one model the host has selected — no client-side model download for hosted chat
Trust model: trusted LAN only (no TLS/pairing in this release). Runtime admission follows commons platformPolicy(); compile-time #if os guards only gate Apple APIs that do not exist on the other Apple platforms
React Native, Flutter, Web, and Electron are not Connect participants in this release

See the iOS/macOS example under examples/ios/RunAnywhereAI/ (ConnectHostManagementView, model-selection Connect section).

Documentation and Examples

Resource	Link
Docs	docs.runanywhere.ai/swift/introduction
Example app	`examples/ios/RunAnywhereAI/`
Contributing / local build	docs/DEVELOPMENT.md

Privacy

Inference runs on-device after models are downloaded. Network access is used for SDK authentication, model downloads, and optional analytics — prompts, responses, and audio are not sent to the cloud for inference by default.

Compliance with regulations such as HIPAA or GDPR depends on your deployment, data handling practices, and backend configuration. Review your architecture and consult legal counsel for regulated use cases.

Support

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

License

This SDK is distributed under the RunAnywhere License. For commercial licensing inquiries, contact [email protected].