Best Practices

Import the api package

import com.runanywhere.sdk.public.RunAnywhere
import com.runanywhere.sdk.public.api.*

The namespaces are extension properties. Without the second import, RunAnywhere.llm does not resolve, and the error message points at the property rather than the missing import.

Initialize in Application.onCreate

class App : Application() {
    override fun onCreate() {
        super.onCreate()
        RunAnywhere.initialize(apiKey = "", baseUrl = "")
    }
}

Phase 2 runs in the background, so initializing at launch gives it time to finish before anything needs it.

Let generation load the model

val result = RunAnywhere.llm.generate(prompt, LlmOptions(model = "qwen3-0.6b"))

Generation auto-loads and downloads when needed. Reach for models.load when you want to control when the cost is paid, such as warming a model behind a splash screen.

Collect in viewModelScope

viewModelScope.launch {
    RunAnywhere.llm.generateStream(prompt).collect { /* … */ }
}

The flow is cancelled when the ViewModel clears, so a generation cannot outlive the screen it belongs to.

Stream anything a person waits for

A 200-token reply takes seconds. Streaming turns that into words appearing immediately.

Cap output length

maxOutputTokens is the biggest lever on latency and battery.

Use	Reasonable cap
Classification, extraction	20 to 50
Chat reply	150 to 300
Spoken reply	100
Summary	150

Unload what you are not using

RunAnywhere.models.unload("qwen3-0.6b")
RunAnywhere.models.unloadAll(ModelCategory.MODEL_CATEGORY_LANGUAGE)

Holding a language model, a speech model, and a vision model at once on a phone is how you get ERROR_CODE_INSUFFICIENT_MEMORY.

Trim conversations

History grows until it exceeds the context window. Keep the system message and the recent turns.

Never swallow CancellationException

catch (e: CancellationException) {
    throw e
}

Catching it breaks structured concurrency: the coroutine keeps running after its scope is gone.

Close what holds hardware

Voice sessions hold the microphone; AudioRecord and AudioTrack hold audio devices. Close them in onCleared, or in a finally.

Use structured output rather than asking nicely

Asking for JSON in a system prompt works most of the time, which is the problem.

RunAnywhere.llm.generateStructured(prompt, schema, StructuredOutputMode.REPAIR)

CONSTRAINED is not implemented and throws.

Test on an ARM64 device

The NPU backend is arm64 only and absent from x86_64 emulators. Performance measured on an emulator means nothing.

Keep the Hugging Face token out of source

RunAnywhere.setHfToken(tokenFromEncryptedPrefs)

Hold it in EncryptedSharedPreferences or the Keystore, never in code, assets, or logs.

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Electron SDK

Vibe Coding

Import the api package

Initialize in Application.onCreate

Let generation load the model

Collect in viewModelScope

Stream anything a person waits for

Cap output length

Unload what you are not using

Trim conversations

Never swallow CancellationException

Close what holds hardware

Use structured output rather than asking nicely

Test on an ARM64 device

Keep the Hugging Face token out of source

​Import the api package

​Initialize in Application.onCreate

​Let generation load the model

​Collect in viewModelScope

​Stream anything a person waits for

​Cap output length

​Unload what you are not using

​Trim conversations

​Never swallow CancellationException

​Close what holds hardware

​Use structured output rather than asking nicely

​Test on an ARM64 device

​Keep the Hugging Face token out of source

Import the api package

Initialize in Application.onCreate

Let generation load the model

Collect in viewModelScope

Stream anything a person waits for

Cap output length

Unload what you are not using

Trim conversations

Never swallow CancellationException

Close what holds hardware

Use structured output rather than asking nicely

Test on an ARM64 device

Keep the Hugging Face token out of source