Onde Inference Kotlin SDK

Run LLMs on-device from Kotlin with Onde Inference. No cloud, no API key, and no user data leaving the device.

Rust SDK · Swift SDK · Flutter SDK · React Native SDK · Website

---

#What is Onde?

Onde is an on-device LLM inference SDK for Kotlin apps. It wraps mistral.rs in a Kotlin-friendly API, with model downloads from Hugging Face, local cache management, and native inference under the hood.

• Runs locally: no cloud, no API key, no user data leaving the device
• Kotlin-friendly: suspend functions and Flow<StreamChunk> for streaming
• Rust underneath: powered by uniffi-rs
• Published on Maven Central: add one Gradle dependency and you're done

This package is built as a Kotlin Multiplatform library. Today that means Android and JVM support, with the same Rust core underneath both. For on-device AI inference for Kotlin and Android, the SDK page is the quickest way to check package docs and platform scope. If you want to test model downloads or fine-tune outputs before moving into app code, use Onde CLI.

---

#Installation

Add to your app's build.gradle.kts:

dependencies {
    implementation("com.ondeinference:onde-inference:0.1.3")
}

Add INTERNET permission to AndroidManifest.xml (required for the initial model download):

<uses-permission android:name="android.permission.INTERNET" />

Minimum requirements - Android 8.0 (API 26) or higher - About 1.1 GB of free storage for the default model (Qwen 2.5 1.5B Q4_K_M)

---

#Quick start on Android

import com.ondeinference.onde.OndeInference
import kotlinx.coroutines.launch

class MainActivity : AppCompatActivity() {

    private val onde by lazy { OndeInference(applicationContext) }

    override fun onCreate(savedInstanceState: Bundle?) {
        super.onCreate(savedInstanceState)

        lifecycleScope.launch {
            // 1. Load the default model (Qwen 2.5 1.5B on Android, ~941 MB)
            //    It downloads from Hugging Face the first time you run it.
            val elapsed = onde.loadDefaultModel(
                systemPrompt = "You are a helpful, concise assistant."
            )
            println("Model loaded in ${elapsed}s")

            // 2. Chat
            val result = onde.chat("What is the capital of Sweden?")
            println(result.text) // → "The capital of Sweden is Stockholm."

            // 3. Clean up
            onde.unload()
        }
    }

    override fun onDestroy() {
        super.onDestroy()
        onde.close()
    }
}

---

#Streaming

If you want tokens as they arrive, use Flow:

lifecycleScope.launch {
    onde.stream("Write a haiku about the ocean.").collect { chunk ->
        // Append each token delta to your UI
        textView.append(chunk.delta)
        if (chunk.done) println("\n[done]")
    }
}

---

#Multi-turn conversation

Onde keeps conversation history for you:

lifecycleScope.launch {
    onde.loadDefaultModel(systemPrompt = "You are a Rust tutor.")

    val r1 = onde.chat("What is ownership?")
    println(r1.text)

    val r2 = onde.chat("Can you give me a code example?")
    println(r2.text) // knows the context from r1

    // Clear history without unloading the model
    onde.clearHistory()
}

---

#Load a specific model

import com.ondeinference.onde.OndeModels

lifecycleScope.launch {
    // Qwen 2.5 1.5B, the default on Android
    val elapsed = onde.loadModel(
        config = OndeModels.qwen25_1_5b(),
        systemPrompt = "You are a coding assistant."
    )
}

---

#Sampling configuration

import com.ondeinference.onde.OndeSampling

lifecycleScope.launch {
    // Default chat settings: temperature 0.7, top_p 0.95, max 512 tokens
    onde.loadDefaultModel(sampling = OndeSampling.default())

    // Deterministic output, useful for factual answers and code
    onde.loadDefaultModel(sampling = OndeSampling.deterministic())

    // Shorter responses, usually better for lower-end devices
    onde.loadDefaultModel(sampling = OndeSampling.mobile())
}

---

#One-shot generation (without changing chat history)

import com.ondeinference.onde.OndeMessage

lifecycleScope.launch {
    val result = onde.generate(
        messages = listOf(
            OndeMessage.system("You are a summariser."),
            OndeMessage.user("Summarise this text: $longArticle"),
        ),
        sampling = OndeSampling.deterministic()
    )
    println(result.text)
}

---

#Engine status

lifecycleScope.launch {
    val info = onde.info()
    println("Status: ${info.status}")          // Ready / Loading / Unloaded / Error
    println("Model: ${info.modelName}")
    println("Memory: ${info.approxMemory}")
    println("History: ${info.historyLength} turns")
}

---

#File system layout

Onde stores model files in the app's internal filesDir. You do not need external storage permissions. INTERNET is only needed for the initial download.

<filesDir>/
├── models/
│   └── hub/
│       └── models--bartowski--Qwen2.5-1.5B-Instruct-GGUF/
│           └── ...
└── tmp/

You can inspect the cache location at runtime:

println(onde.modelCacheDir.absolutePath)

---

#How it fits together

┌───────────────────────────────────────┐
│  Kotlin app / Android activity        │
│  import com.ondeinference.onde.*      │
└──────────────┬────────────────────────┘
               │ suspend / Flow
               ▼
┌───────────────────────────────────────┐
│  OndeInference.kt  (this SDK)         │
│  Android filesystem setup             │
│  Dispatchers.IO dispatch              │
└──────────────┬────────────────────────┘
               │ UniFFI-generated Kotlin (onde.kt)
               ▼
┌───────────────────────────────────────┐
│  libonde.so  (Rust, per ABI)          │
│  mistral.rs  +  Candle  (CPU)         │
│  HuggingFace Hub download             │
└───────────────────────────────────────┘

On Android, the Rust engine is compiled for four ABIs and bundled as JNI libs:

ABI	Architecture
arm64-v8a	64-bit ARM (modern Android phones)
armeabi-v7a	32-bit ARM (older phones)
x86_64	Android emulators (Intel/AMD)
x86	32-bit emulators

---

#Building from source

#Prerequisites

# Rust Android targets
rustup target add aarch64-linux-android armv7-linux-androideabi x86_64-linux-android i686-linux-android

# cargo-ndk — manages NDK toolchain for cross-compilation
cargo install cargo-ndk

# Android NDK (via Android Studio or SDK manager)
export ANDROID_NDK_HOME=$HOME/Library/Android/sdk/ndk/26.1.10909125

#Build the native libraries

# From the repo root
./sdk/kotlin/scripts/build-android.sh

This produces .so files in sdk/kotlin/lib/src/androidMain/jniLibs/.

#Generate Kotlin UniFFI bindings

./sdk/kotlin/scripts/generate-bindings.sh

This generates onde.kt in sdk/kotlin/lib/src/generated/kotlin/. The generated file is gitignored, so regenerate it whenever the Rust API changes.

#Build the Android artifact

cd sdk/kotlin
./gradlew :lib:assembleRelease
# Output: lib/build/outputs/aar/lib-release.aar

---

#Publishing to Maven Central

Publishing is handled by CI. It requires:

Secret	Description
MAVEN_CENTRAL_USERNAME	Sonatype Central Portal username
MAVEN_CENTRAL_PASSWORD	Sonatype Central Portal password
SIGNING_KEY_ID	PGP key ID (last 8 hex chars)
SIGNING_KEY	ASCII-armored PGP private key
SIGNING_PASSWORD	PGP key passphrase

Tag a release to trigger the workflow:

# Bump version in Cargo.toml and gradle.properties first
git tag 0.1.4 && git push origin 0.1.4

The CI workflow builds the native libraries, generates bindings, packages the Android artifact, signs it, and publishes to Maven Central automatically.

---

#License

MIT OR Apache-2.0. See LICENSE.

#Copyright

© 2026 Onde Inference (Splitfire AB).