Skip to main content

​
Overview

The RunAnywhere Kotlin SDK runs AI models on the Android device itself. Once a model is downloaded, inference needs no network, and the audio, text, and images it processes stay on the phone.A single C++ core implements the model lifecycle, inference routing, and pipeline orchestration. The Kotlin SDK is a typed wrapper over it, so behavior matches the Swift, Flutter, React Native, and Web SDKs rather than being reimplemented per platform.

LLM

Text generation with streaming support via Kotlin Flows

VLM

Vision language models for image understanding

LoRA

Hot-swap fine-tuned adapters at runtime

RAG

On-device retrieval-augmented generation

STT

Speech-to-text transcription with Whisper models

TTS

Neural voice synthesis with Sherpa-ONNX

Tool Calling

Function calling with typed tool definitions

VAD

Real-time voice activity detection

Voice Agent

Full VAD → STT → LLM → TTS pipeline

Hexagon NPU

Qualcomm QHexRT backend for LLM, VLM, STT, TTS, embeddings, and diffusion

​
Key capabilities

Backends are separate modules: llama.cpp for GGUF language and vision models, ONNX Runtime with Sherpa-ONNX for speech, and QHexRT for the Qualcomm Hexagon NPU. Link only the ones you use.The public surface is Kotlin coroutines and Flows throughout. Every structured type is generated from the same Protobuf IDL the C++ core is generated from, so field names and defaults are identical across the Swift, Kotlin, Flutter, React Native, and Web SDKs.

​
Core philosophy

Inference runs locally. Once a model is downloaded, no network connection is needed to use it.
Each backend is a separate artifact carrying its own native libraries. Link only the ones you use, so an LLM-only app does not ship the speech stack.
Audio, text, and images are processed on the device unless you explicitly configure a cloud route.
Business logic lives in the shared C++ core and the structured types are generated from one Protobuf IDL, so the Kotlin API mirrors Swift, Flutter, React Native, and Web.

​
Features

Every capability hangs off the RunAnywhere object as a namespace. The namespaces are extension properties in com.runanywhere.sdk.public.api, so importing com.runanywhere.sdk.public.api.llm gives you RunAnywhere.llm.One-shot verbs are suspend functions and streaming verbs return a Flow. Every stream uses the same grammar: a started event, then deltas, then a completed event carrying the aggregate result. Cancelling the Flow cancels that one request.Reasoning models are handled by the SDK rather than by prompt text. LlmOptions.reasoning decides whether the model thinks and whether those tokens reach you, and every streamed token carries a TokenKind of TEXT or THOUGHT.

​
System requirements

ARM64 is the primary target. The Hexagon NPU module is arm64-v8a only.

​
SDK modules

RAG is part of the core SDK and uses embedding models served by the ONNX module. There is no separate RAG artifact. See Installation for versions and sizes.

​
Architecture

​
Next steps

Installation

Add the SDK to your project via Gradle

Quick start

Initialize the SDK and run your first inference