react-native-litert-lm

High-performance on-device LLM inference for React Native, powered by LiteRT-LM and Nitro Modules. Optimized for Gemma 4 and other on-device language models.

Features

• 🚀 Native Performance — Kotlin (Android) / C++ (iOS) via Nitro Modules JSI bindings
• 🧠 Gemma 4 Ready — First-class support for Gemma 4 E2B/E4B multimodal models (text + vision + audio)
• ⚡ GPU Acceleration — GPU delegate (Android), Metal/MPS (iOS)
• 🔄 Streaming Support — Token-by-token generation callbacks
• 📱 Cross-Platform — Android API 26+ / iOS 15.0+
• 🖼️ Multimodal — Image and audio input support
• 🧵 Async API — Non-blocking inference on dedicated large-stack threads
• 📊 Real Memory Tracking — OS-level memory metrics (RSS, native heap, available memory) via native APIs
• 🧮 Zero-Copy Buffers — Memory snapshots stored in native ArrayBuffers via Nitro Modules
• 📥 Automatic Model Download — Downloads models from URL with progress tracking and local caching

Installation

npm install react-native-litert-lm react-native-nitro-modules

Expo

Add to your app.json:

{
  "expo": {
    "plugins": ["react-native-litert-lm"],
    "android": {
      "minSdkVersion": 26
    }
  }
}

Then create a development build:

npx expo prebuild
npx expo run:android  # Android
npx expo run:ios      # iOS

Note: Only ARM devices/simulators are supported. x86_64 Android emulators are not supported.

Bare React Native

# Android
cd android && ./gradlew clean

# iOS
cd ios && pod install

Example App

The example/ directory contains a fully functional test app with a dark-themed diagnostic UI that demonstrates:

• Model downloading with progress tracking
• Text inference (blocking and streaming)
• Multi-turn conversation with context retention
• Performance benchmarking (tokens/sec, latency)
• Real-time memory tracking
• Quick chat interface

Running the Example

1. Build the library (compiles TypeScript to lib/):

   npm run build

2. Install example dependencies:

   cd example
   npm install

3. Create a development build and run:

   npx expo prebuild --clean
   npx expo run:android  # Android
   npx expo run:ios      # iOS (requires XCFramework — see "Building the iOS Engine" below)

Note: If you change native code (C++/Kotlin/Obj-C++), you must run npx expo prebuild --clean again before rebuilding.

Model Management

LiteRT-LM models (like Gemma 4) are large files (2–4 GB) and cannot be bundled into your app binary. They are downloaded at runtime.

Automatic Downloading

The library handles downloading automatically when you pass a URL to loadModel or useModel. Downloads include:

• Progress tracking — real-time download percentage via callbacks
• Local caching — downloaded models are cached and reused across app launches
  • Android: app-local temp directory
  • iOS: Library/Caches/litert_models/ (survives app relaunch; reclaimable by iOS under storage pressure)
• HTTPS enforcement — only secure URLs are accepted

Manual Downloading (Optional)

If you prefer to manage downloads yourself (e.g., using expo-file-system), download the .litertlm file to a local path and pass that path to the library:

import * as FileSystem from "expo-file-system";
import { GEMMA_4_E2B_IT } from "react-native-litert-lm";

const localPath = `${FileSystem.documentDirectory}gemma-4-E2B-it.litertlm`;

async function downloadModel() {
  const info = await FileSystem.getInfoAsync(localPath);
  if (info.exists) return localPath;

  await FileSystem.downloadAsync(GEMMA_4_E2B_IT, localPath);
  return localPath;
}

Usage

React Hook (Recommended)

The useModel hook manages the full model lifecycle: downloading, loading, inference, and cleanup.

import { useModel, GEMMA_4_E2B_IT } from "react-native-litert-lm";
import { Platform } from "react-native";

function App() {
  const {
    model,
    isReady,
    downloadProgress,
    error,
    load,          // Manually trigger load
    deleteModel,   // Delete cached model file
    memorySummary, // Auto-updated memory stats (if tracking enabled)
  } = useModel(GEMMA_4_E2B_IT, {
    backend: 'cpu',
    autoLoad: true, // Default: true. Set false to load manually via load().
    systemPrompt: "You are a helpful assistant.",
    enableMemoryTracking: true,
  });

  if (!isReady) {
    return <Text>Loading... {Math.round(downloadProgress * 100)}%</Text>;
  }

  const generate = async () => {
    const response = await model.sendMessage("Hello!");
    console.log(response);
  };

  return <Button title="Generate" onPress={generate} />;
}

Manual Usage

import { createLLM } from "react-native-litert-lm";

const llm = createLLM();

// Load a model from URL (auto-downloads) or local path
await llm.loadModel("https://example.com/model.litertlm", {
  backend: "gpu",
  systemPrompt: "You are a helpful assistant.",
});

// Generate a response
const response = await llm.sendMessage("What is the capital of France?");
console.log(response);

// Clean up
llm.close();

Streaming Generation

llm.sendMessageAsync("Tell me a story", (token, done) => {
  process.stdout.write(token);
  if (done) console.log("\n--- Done ---");
});

Multimodal (Image / Audio)

Note: Multimodal is fully supported on Android. iOS has the code paths implemented but vision/audio executors may not be available in the current XCFramework build — use checkMultimodalSupport() to verify at runtime.

import { checkMultimodalSupport } from "react-native-litert-lm";

const warning = checkMultimodalSupport();
if (warning) {
  console.warn(warning); // Experimental on iOS
} else {
  // Image input (for vision models like Gemma 4)
  // Images >1024px are automatically resized to prevent OOM
  const response = await llm.sendMessageWithImage(
    "What's in this image?",
    "/path/to/image.jpg",
  );

  // Audio input
  const transcription = await llm.sendMessageWithAudio(
    "Transcribe this audio",
    "/path/to/audio.wav",
  );
}

Performance Stats

const stats = llm.getStats();
console.log(`Generated ${stats.completionTokens} tokens`);
console.log(`Speed: ${stats.tokensPerSecond.toFixed(1)} tokens/sec`);
console.log(`Time to first token: ${stats.timeToFirstToken.toFixed(0)} ms`);

Memory Tracking

The library provides real OS-level memory data — no estimation. It reads directly from mach_task_basic_info (iOS) and Debug.getNativeHeapAllocatedSize() + /proc/self/status (Android).

Direct Memory Query

const usage = llm.getMemoryUsage();
console.log(
  `Native heap: ${(usage.nativeHeapBytes / 1024 / 1024).toFixed(1)} MB`,
);
console.log(`RSS: ${(usage.residentBytes / 1024 / 1024).toFixed(1)} MB`);
console.log(
  `Available: ${(usage.availableMemoryBytes / 1024 / 1024).toFixed(1)} MB`,
);
console.log(`Low memory: ${usage.isLowMemory}`);

Automatic Tracking with Native Buffers

Enable memory tracking to automatically record snapshots in a native-backed ArrayBuffer after every inference call:

const llm = createLLM({
  enableMemoryTracking: true,
  maxMemorySnapshots: 256,
});

await llm.loadModel("/path/to/model.litertlm", { backend: "cpu" });
await llm.sendMessage("Hello!");

const summary = llm.memoryTracker!.getSummary();
console.log(
  `Peak RSS: ${(summary.peakResidentBytes / 1024 / 1024).toFixed(1)} MB`,
);
console.log(
  `RSS Delta: ${(summary.residentDeltaBytes / 1024 / 1024).toFixed(1)} MB`,
);

Using useModel with Memory Tracking

const { model, isReady, memorySummary } = useModel(modelUrl, {
  enableMemoryTracking: true,
  maxMemorySnapshots: 100,
});

// memorySummary auto-updates after each inference call
if (memorySummary) {
  console.log(`Current RSS: ${memorySummary.currentResidentBytes}`);
  console.log(`Peak RSS: ${memorySummary.peakResidentBytes}`);
}

Standalone Memory Tracker

import {
  createMemoryTracker,
  createNativeBuffer,
} from "react-native-litert-lm";

const tracker = createMemoryTracker(100);

tracker.record({
  timestamp: Date.now(),
  nativeHeapBytes: 50_000_000,
  residentBytes: 200_000_000,
  availableMemoryBytes: 4_000_000_000,
});

// Access the underlying native buffer (zero-copy transfer to native code)
const buffer = tracker.getNativeBuffer();

Supported Models

Download .litertlm models automatically using the exported URL constants, or manually from HuggingFace:

| Constant | Model | Size | Min RAM | Auth Required | | :--------------------- | :--------------------------------- | :------- | :------ | :------------ | | GEMMA_4_E2B_IT | Gemma 4 E2B (Multimodal, IT) | 2.58 GB | 4 GB+ | ❌ No | | GEMMA_4_E4B_IT | Gemma 4 E4B (Higher Quality) | 3.65 GB | 6 GB+ | ❌ No | | GEMMA_3N_E2B_IT_INT4 | Gemma 3n E2B (Int4, Multimodal) | ~1.3 GB | 4 GB+ | ✅ HuggingFace |

Recommended: Use GEMMA_4_E2B_IT for most use cases. It's multimodal (text + vision + audio) and downloads directly from HuggingFace without requiring an account.

iOS Note: Models larger than ~2 GB (like Gemma 4) require the com.apple.developer.kernel.extended-virtual-addressing entitlement. See iOS Entitlements below.

Other compatible models (download manually from HuggingFace):

| Model | Size | Min RAM | Notes | | ------------- | ------- | ------- | --------------------- | | Gemma 3 1B | ~1 GB | 4 GB+ | Smallest, fastest | | Phi-4 Mini | ~2 GB | 4 GB+ | Microsoft's small LLM | | Qwen 2.5 1.5B | ~1.5 GB | 4 GB+ | Multilingual |

API Reference

createLLM(options?): LiteRTLM

Creates a new LLM inference engine instance.

• options.enableMemoryTracking — enable automatic memory snapshot recording
• options.maxMemorySnapshots — max number of snapshots to retain (default: 256)

loadModel(path, config?): Promise<void>

Loads a model from a local path or HTTPS URL.

| Parameter | Type | Default | Description | | --------------------- | -------- | ------- | ----------------------------------------- | | path | string | — | Absolute path to .litertlm or HTTPS URL | | config.backend | string | 'cpu' | 'cpu', 'gpu', or 'npu' | | config.systemPrompt | string | — | System prompt for the model | | config.temperature | number | 0.7 | Sampling temperature | | config.topK | number | 40 | Top-K sampling | | config.topP | number | 0.95 | Top-P (nucleus) sampling | | config.maxTokens | number | 1024 | Maximum generation length |

Backend Options

| Backend | Engine | Speed | Notes | | ------- | ------------------- | ------- | ---------------------------------------------- | | 'cpu' | CPU inference | Slowest | Always available, lower RAM requirement | | 'gpu' | GPU / Metal | Fast | Recommended default | | 'npu' | NPU / Neural Engine | Fastest | Requires supported hardware; falls back to GPU |

iOS: 'cpu' is the recommended default backend. 'gpu' (Metal/MPS) is also supported. The engine automatically tries multiple backend combinations if the primary one fails.

sendMessage(message): Promise<string>

Runs inference synchronously on a background thread. Returns the complete response.

sendMessageAsync(message, callback)

Streaming generation. Callback signature: (token: string, isDone: boolean) => void.

sendMessageWithImage(message, imagePath): Promise<string>

Send a message with an image (for vision models like Gemma 4 E2B).

sendMessageWithAudio(message, audioPath): Promise<string>

Send a message with audio (for audio-capable models like Gemma 4 E2B).

getStats(): GenerationStats

Returns performance metrics from the last inference call.

interface GenerationStats {
  tokensPerSecond: number;
  totalTime: number; // seconds
  timeToFirstToken: number; // seconds
  promptTokens: number;
  completionTokens: number;
  prefillSpeed: number; // tokens/sec
}

getMemoryUsage(): MemoryUsage

Returns real OS-level memory usage.

interface MemoryUsage {
  nativeHeapBytes: number;
  residentBytes: number;
  availableMemoryBytes: number;
  isLowMemory: boolean;
}

getHistory(): Message[]

Returns the conversation history.

resetConversation()

Clears conversation context and starts a fresh session.

close()

Releases all native resources. Call when the model is no longer needed.

deleteModel(fileName): Promise<void>

Deletes a cached model file from the app's local storage.

Utility Functions

import {
  checkBackendSupport,
  checkMultimodalSupport,
  getRecommendedBackend,
  applyGemmaTemplate,
  applyPhiTemplate,
  applyLlamaTemplate,
} from "react-native-litert-lm";

// Check if a backend is supported
const warning = checkBackendSupport("npu"); // string | undefined
const mmError = checkMultimodalSupport(); // string | undefined
const backend = getRecommendedBackend(); // 'gpu' | 'cpu'

// Manual prompt formatting (advanced)
const prompt = applyGemmaTemplate(
  [{ role: "user", content: "Hello!" }],
  "You are helpful.",
);

Requirements

| Dependency | Version | | -------------------------- | ------------- | | React Native | 0.76+ | | react-native-nitro-modules | 0.35.0+ | | Android API | 26+ (ARM64) | | iOS | 15.0+ (ARM64) | | LiteRT-LM Engine | 0.10.1 |

Platform Support

| Platform | Status | Architecture | Backends | | -------- | -------- | ------------ | ---------------- | | Android | ✅ Ready | arm64-v8a | CPU, GPU, NPU | | iOS | ✅ Ready | arm64 | CPU, GPU (Metal) |

iOS Feature Matrix

| Feature | Status | Notes | | ---------------------------- | ------ | ----------------------------------------------------- | | Text inference (blocking) | ✅ | Via LiteRT-LM C API | | Text inference (streaming) | ✅ | Token-by-token callbacks | | CPU inference | ✅ | Recommended default backend | | GPU inference (Metal/MPS) | ✅ | Supported via backend: 'gpu' | | Model download with progress | ✅ | NSURLSession, cached in Caches/ | | Memory tracking | ✅ | mach_task_basic_info | | Multi-turn conversation | ✅ | Context retained across turns | | Multimodal (image/audio) | 🧪 | Code paths exist; vision/audio executors experimental | | Constrained decoding | ❌ | Requires llguidance Rust runtime | | Function calling | ❌ | Requires Rust CXX bridge runtime |

iOS Entitlements

Models larger than ~2 GB (like Gemma 4 E2B at 2.58 GB) require the Extended Virtual Addressing entitlement on iOS physical devices. Without it, iOS limits virtual memory to ~2 GB and the app will be killed by Jetsam.

Add to your app's .entitlements file:

<key>com.apple.developer.kernel.extended-virtual-addressing</key>
<true/>

Note: This entitlement requires a paid Apple Developer account ($99/year). Gemma 3n E2B (~1.3 GB) works without it.

Building the iOS Engine

The iOS build uses a Bazel-to-XCFramework pipeline that compiles the LiteRT-LM C engine and all transitive dependencies into a static library (~83 MB).

Prerequisites

• Bazel 7.6.1+ (via Bazelisk recommended)
• Xcode command line tools (xcode-select --install)

Build

./scripts/build-ios-engine.sh

This will:

1. Clone/checkout LiteRT-LM v0.10.1 source into .litert-lm-build/
2. Build //c:engine for ios_arm64 and ios_sim_arm64 via Bazel
3. Collect all transitive .o files (engine, protobuf, re2, sentencepiece, etc.)
4. Compile C/C++ stubs for unavailable Rust dependencies
5. Patch PromptTemplate to use a simplified template engine (no Rust MinijinjaTemplate)
6. Merge ~1,900 object files into a static library via libtool
7. Package into ios/Frameworks/LiteRTLM.xcframework

Output

ios/Frameworks/LiteRTLM.xcframework/
├── Info.plist
├── ios-arm64/LiteRTLM.framework/              # Device
│   ├── LiteRTLM                                # ~81 MB static library
│   └── Headers/litert_lm_engine.h
└── ios-arm64-simulator/LiteRTLM.framework/    # Simulator
    ├── LiteRTLM                                # ~83 MB static library
    └── Headers/litert_lm_engine.h

FFI Stubs

Certain LiteRT-LM features depend on Rust libraries (llguidance, CXX bridge, MinijinjaTemplate) that are not available in the iOS Bazel build. These are replaced with stubs:

| Stub File | Location | Purpose | | ------------------------------------ | ---------------- | ---------------------------------------- | | cxx_bridge_stubs.cc | scripts/stubs/ | CXX bridge runtime + Rust FFI type stubs | | llguidance_stubs.c | scripts/stubs/ | llguidance constrained decoding C API | | gemma_model_constraint_provider.cc | scripts/stubs/ | Gemma constraint provider factory |

Additionally, PromptTemplate is patched at build time to use a simplified C++ template formatter instead of the Rust MinijinjaTemplate, which avoids all Rust FFI calls during conversation setup.

Text inference works fully without these Rust components. Only constrained decoding, function calling parsers, and advanced Jinja2 template features are affected.

Architecture

┌─────────────────────────────────────────────────┐
│  React Native (TypeScript)                      │
│  useModel() / createLLM() / sendMessage()       │
├─────────────────────────────────────────────────┤
│  Nitro Modules JSI Bridge                       │
├──────────────────────┬──────────────────────────┤
│  Android (Kotlin)    │  iOS (C++)               │
│  HybridLiteRTLM.kt   │  HybridLiteRTLM.cpp      │
│  litertlm-android    │  LiteRTLM C API          │
│  AAR (GPU delegate)  │  XCFramework (Metal)     │
└──────────────────────┴──────────────────────────┘

• Android: Kotlin (HybridLiteRTLM.kt) interfacing with the litertlm-android AAR.
• iOS: C++ (HybridLiteRTLM.cpp) interfacing with the LiteRT-LM C API via a prebuilt LiteRTLM.xcframework. All engine operations (load, inference, streaming) run on dedicated pthread threads with 8 MB stack to accommodate XNNPack's stack requirements. Platform-specific code (model downloading, file management) is in Objective-C++ (ios/IOSDownloadHelper.mm).

For contributors: Changes to cpp/HybridLiteRTLM.cpp do not affect Android. Feature changes must be applied to both the Kotlin and C++ implementations.

License

The code in this repository is licensed under the MIT License.

⚠️ AI Model Disclaimer

This library is an execution engine for on-device LLMs. The AI models themselves are not distributed with this package and have their own licenses:

• Gemma (Google): Gemma Terms of Use
• Llama 3 (Meta): Llama 3.2 Community License
• Qwen (Alibaba): Apache 2.0
• Phi (Microsoft): MIT License

By downloading and using these models, you agree to their respective licenses and acceptable use policies. The author of react-native-litert-lm takes no responsibility for model outputs or applications built with them.