react-native-nitro-tflite

High-performance TensorFlow Lite library for React Native, powered by Nitro Modules.

Note: This is an unofficial Nitro Module migration of react-native-fast-tflite, fully optimized for the New Architecture, Bridgeless mode, and Vision Camera frame processors.

Features

• ✅ HybridObject-based — No install() or global JSI functions
• ✅ Works from any frame processor / worklet — Callable from any thread or runtime
• ✅ Bridge / Bridgeless agnostic — Works on both architectures automatically
• ✅ Same API — Drop-in replacement for react-native-fast-tflite
• ✅ Model caching — Identical calls return the same model with zero re-allocation
• ✅ Progress callbacks — Track large model downloads in real time
• ✅ Dynamic input shapes — Resize input tensors at runtime
• ✅ Metal GPU delegate (iOS) — Hardware-accelerated inference via Apple Metal

---

Installation

npm install react-native-nitro-tflite react-native-nitro-modules

iOS

cd ios && pod install

Android

No additional steps needed — TFLite libraries are downloaded automatically via Gradle.

---

Usage

Basic — React hook

import { useTensorflowModel } from 'react-native-nitro-tflite'

function App() {
  const model = useTensorflowModel(require('./model.tflite'))

  if (model.state === 'loaded') {
    const output = model.model.runSync([inputFloat32Array])
    console.log('Output:', output)
  }

  if (model.state === 'error') {
    console.error('Failed to load:', model.error)
  }

  return <View />
}

Load imperatively

import { loadTensorflowModel } from 'react-native-nitro-tflite'

const model = await loadTensorflowModel(require('./model.tflite'))
const output = model.runSync([inputData])

Load from URL

const model = await loadTensorflowModel({ url: 'https://example.com/model.tflite' })

---

Progress Callback

Track download progress for large models (useful for remote URLs):

import { loadTensorflowModel } from 'react-native-nitro-tflite'

const model = await loadTensorflowModel(
  { url: 'https://example.com/large_model.tflite' },
  'default',
  (progress) => {
    if (progress === -1) {
      console.log('Downloading… (size unknown)')
    } else {
      console.log(`Downloading: ${Math.round(progress * 100)}%`)
    }
  }
)

progress is in [0, 1]. Receives -1 when the server does not send a Content-Length header.

---

Model Caching

Models are automatically cached in native memory by URL + delegate.
Calling loadTensorflowModel twice with the same arguments returns the same model instantly:

const m1 = await loadTensorflowModel(require('./model.tflite'))
const m2 = await loadTensorflowModel(require('./model.tflite')) // cache hit — instant
// m1 and m2 share the same underlying TfLiteInterpreter

To eagerly free all cached models (e.g. on screen unmount):

import { clearTensorflowModelCache } from 'react-native-nitro-tflite'

clearTensorflowModelCache()

---

Dynamic Input Shape

For models that support dynamic tensor shapes, call reshapeInput before inference:

const model = await loadTensorflowModel(require('./dynamic_model.tflite'))

// Switch to 640×640 input at runtime
model.reshapeInput(0, [1, 640, 640, 3])

const output = model.runSync([new Float32Array(1 * 640 * 640 * 3)])

After reshapeInput, both input and output tensor sizes are re-allocated automatically.

---

Delegates

Delegate	Platform	Description	Setup
default	Both	CPU inference (always available)	—
core-ml	iOS	Apple CoreML acceleration	$EnableCoreMLDelegate = true in Podfile
metal	iOS	Metal GPU acceleration	$EnableMetalDelegate = true in Podfile
nnapi	Android	Android Neural Networks API	—
android-gpu	Android	Android GPU delegate	—

Enabling CoreML (iOS)

# ios/Podfile
$EnableCoreMLDelegate = true

use_frameworks! ...

Enabling Metal (iOS)

# ios/Podfile
$EnableMetalDelegate = true

use_frameworks! ...

Then re-run pod install.

---

Metro Configuration

Add tflite as an asset extension in metro.config.js:

const { getDefaultConfig } = require('@react-native/metro-config')

const config = getDefaultConfig(__dirname)
config.resolver.assetExts.push('tflite')

module.exports = config

---

Architecture

This library uses Nitro Modules with manual C++ HybridObject implementation:

• HybridTfliteModelFactory — Singleton factory; loads models on a background thread, manages the cache, handles delegate selection
• HybridTfliteModel — Wraps TfLiteInterpreter; exposes run / runSync / reshapeInput / inputs / outputs / delegate
• TfliteModelHostObject — A jsi::HostObject wrapper that makes the model usable across worklet runtimes (VisionCamera frame processors)

JS:  loadTensorflowModel(source, delegate, onProgress?)
       ↓
     resolveSourceUri()  →  URI string
       ↓
     factory.loadModel(uri, delegate, onProgress)
       ↓  [cache hit → resolve immediately]
       ↓  [cache miss → background thread]
C++: HybridTfliteModelFactory::loadModelRaw()
       ↓
     fetchURL(uri, onProgress)   ← platform-specific (iOS: NSURLSession, Android: OkHttp)
       ↓
     TfLiteModelCreate() + TfLiteInterpreterCreate()
       ↓
     HybridTfliteModel  →  TfliteModelHostObject  →  JS Promise resolved
       ↓
     _modelCache[key] = weak_ptr<HybridTfliteModel>

---

License

MIT