Offline AI on Android: TensorFlow Lite with GPU/NNAPI Acceleration

We often encounter this situation: you have a trained object detection model in PyTorch and want to run it on a smartphone without internet. You convert it to TensorFlow Lite, add it to assets—and on a test Pixel 6 it flies. But on a Samsung Galaxy A21s (Exynos) the app crashes with OutOfMemoryError, and on a Xiaomi Redmi Note 8 (Qualcomm) it works but lags. The cause is the choice of acceleration delegate and memory management. TensorFlow Lite is the de facto standard for on-device ML, but its integration requires deep understanding of hardware specifics. Savings on cloud computing can reach 90%—thousands of dollars per month for a production service—but only with proper implementation. Our team has delivered solutions for 40+ projects, with average savings of $1500–$3000 per month per client. For one logistics client, we deployed a package detection model on Android scanners, cutting cloud inference costs from $5,000/month to near zero. The solution paid for itself in under 2 months.

TensorFlow Lite official documentation

How to Convert a Model with Minimal Loss?

The first step is to export the model from PyTorch/ONNX to TensorFlow. Then use TFLiteConverter with optimizations. For INT8, calibrate on a representative dataset. Example:

converter = tf.lite.TFLiteConverter.from_saved_model("model_tf")
converter.optimizations = [tf.lite.Optimize.DEFAULT]  # динамическая квантизация FP16
converter.target_spec.supported_types = [tf.float16]  # для GPU delegate
converter.representative_dataset = representative_dataset  # для INT8 (калибровка)
tflite_model = converter.convert()
with open("model_fp16.tflite", "wb") as f:
    f.write(tflite_model)

Why Delegate Choice Is Critical for Performance?

We use a hybrid configuration: GPU with fallback to XNNPACK and NNAPI as a last resort. Here’s how it looks in code:

import org.tensorflow.lite.gpu.GpuDelegate
import org.tensorflow.lite.gpu.CompatibilityList

val compatList = CompatibilityList()
val options = Interpreter.Options().apply {
    if (compatList.isDelegateSupportedOnThisDevice) {
        addDelegate(GpuDelegate(compatList.bestOptionsForThisDevice))
    } else {
        // Fallback: сначала NNAPI, если не сработает — XNNPACK
        setUseNNAPI(true)
        setUseXNNPACK(true)
    }
    setNumThreads(Runtime.getRuntime().availableProcessors())
}

var interpreter: Interpreter? = null
try {
    interpreter = Interpreter(FileUtil.loadMappedFile(context, "model_fp16.tflite"), options)
    // Тестовый прогон (нужен для выявления ошибок NNAPI)
    interpreter.run(testInput, testOutput)
} catch (e: Exception) {
    Log.w("ML", "NNAPI failed, fallback to CPU: ${e.message}")
    options.setUseNNAPI(false)
    interpreter = Interpreter(modelBuffer, options)
}

NNAPI is unstable in practice: on some devices it gives 5× speedup, on others it crashes. Always wrap the launch in try/catch. Without this, stability is not guaranteed.

Managing Buffers: ByteBuffer vs TensorBuffer

Direct ByteBuffer management is faster but verbose. TensorBuffer from org.tensorflow.lite.support is more convenient and less error-prone:

import org.tensorflow.lite.support.image.ImageProcessor
import org.tensorflow.lite.support.image.TensorImage
import org.tensorflow.lite.support.common.ops.NormalizeOp
import org.tensorflow.lite.support.image.ops.ResizeOp

val imageProcessor = ImageProcessor.Builder()
    .add(ResizeOp(224, 224, ResizeOp.ResizeMethod.BILINEAR))
    .add(NormalizeOp(127.5f, 127.5f))
    .build()

val tensorImage = TensorImage(DataType.FLOAT32)
tensorImage.load(bitmap)
val processedImage = imageProcessor.process(tensorImage)

val outputBuffer = TensorBuffer.createFixedSize(intArrayOf(1, 1000), DataType.FLOAT32)
interpreter.run(processedImage.buffer, outputBuffer.buffer)

val probabilities = outputBuffer.floatArray
val topIndex = probabilities.indices.maxByOrNull { probabilities[it] } ?: -1

CameraX Integration

val imageAnalyzer = ImageAnalysis.Builder()
    .setTargetResolution(Size(640, 480))
    .setBackpressureStrategy(ImageAnalysis.STRATEGY_KEEP_ONLY_LATEST)
    .build()
    .also {
        it.setAnalyzer(cameraExecutor) { imageProxy ->
            try {
                val bitmap = imageProxy.toBitmap()
                runInference(bitmap)
            } finally {
                imageProxy.close()  // КРИТИЧНО: иначе CameraX зависнет
            }
        }
    }

imageProxy.close() in a finally block is not optional. If you don't close the ImageProxy, CameraX stops delivering new frames after a few seconds.

Why Numerical Accuracy After Quantization Matters?

After conversion, always check accuracy on a test set. FP16 typically loses <1%, INT8 loses 1–3%. If losses are larger, the calibration dataset may be too small or the model sensitive to specific layers. Normal maximum deviation is 0.01 for FP16 and 0.05 for INT8. If deviation is higher, return to conversion: change optimizations or replace sensitive layers.

Performance Table on Different Chips (Example)

What the Work Includes

• Model conversion from your framework (PyTorch, ONNX, SavedModel) with optimization selection.
• Delegate selection and setup with fallback logic.
• Integration with CameraX or other data source.
• Numerical accuracy testing and profiling with Android Profiler + TFLite Benchmark Tool.
• Build testing on 10+ physical devices (different chips, Android versions).
• Documentation (API description, architecture), access to the code repository, developer training, and 2 weeks of support.

Typical Mistakes in TFLite Integration

• Forgetting to close ImageProxy—frame stream stops.
• Not checking delegate support on the device—crashes on old chips.
• Using INT8 without calibration—accuracy drop >10%.
• Loading the model into RAM without MappedByteBuffer—OOM on devices with 2 GB RAM.

Time Estimates

Basic TFLite model integration on Android takes 1–2 weeks. With multi-delegate logic, CameraX pipeline, testing on a device park—3–5 weeks. The cost is calculated individually based on complexity. Get a consultation for your task—we'll provide timelines and cost.

We are a team with in-house ML development experience, certified Google engineers. With 40+ projects behind us, including offline AI applications for automotive and medical industries. Contact us for an accurate estimate of your task—we'll send a plan and timeline for free.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.