On-Device ML Integration with ONNX Runtime for Mobile Apps

Imagine a mobile app that must process images, text, or sound without network access. Server latency is unacceptable, data privacy is critical. Every extra megabyte of traffic costs the user. On-device ML solves these issues, and ONNX Runtime is the key tool for cross-platform deployment. We integrate it so that the model runs equally fast on iOS and Android. Switching to on-device can reduce server infrastructure costs by up to 90%, saving hundreds of thousands of rubles monthly under high loads.

ONNX Runtime Mobile appeals with one argument: one model, both platforms. Convert PyTorch or TensorFlow to ONNX, add onnxruntime-android and onnxruntime-objc, run the same .onnx file. In practice, the difference in execution providers between iOS and Android still requires platform-specific code, but the model itself is unified. Our experience: over 5 years in mobile ML, dozens of on-device inference projects. Contact us for an assessment of your model and a preliminary quote.

How to Prepare a Model for Mobile

Standard ONNX export from PyTorch:

import torch
import onnx
from onnxsim import simplify  # onnx-simplifier for graph optimization

model = MyModel(); model.eval()
dummy = torch.zeros(1, 3, 224, 224)

torch.onnx.export(
    model, dummy, "model.onnx",
    opset_version=17,
    input_names=["input"],
    output_names=["output"],
    dynamic_axes={"input": {0: "batch_size"}, "output": {0: "batch_size"}}
)

# Simplify graph — removes redundant reshapes, transposes, makes graph cleaner
model_onnx = onnx.load("model.onnx")
model_simplified, check = simplify(model_onnx)
 onnx.save(model_simplified, "model_simplified.onnx")

Additionally for mobile — quantization via onnxruntime.quantization:

from onnxruntime.quantization import quantize_dynamic, QuantType

quantize_dynamic(
    "model_simplified.onnx",
    "model_int8.onnx",
    weight_type=QuantType.QInt8
)
# Model size reduces ~4× compared to FP32

Which Execution Provider Delivers Maximum Performance?

Android: NNAPI vs XNNPACK

On Android, the choice of Execution Provider depends on hardware. NNAPI delegates operations to NPU/DSP, providing up to 2x acceleration on supported ops. XNNPACK is an optimized CPU backend using SIMD instructions, speeding up to 2x on CPU but without NPU access. On a project with object detection on MediaTek Dimensity, we got 45 ms on NNAPI vs 80 ms on XNNPACK. We recommend using NNAPI for devices with NPU, XNNPACK as fallback.

iOS: CoreML Execution Provider

appendCoreMLExecutionProvider on iOS 13+ delegates supported operations to Core ML, gaining access to ANE. Operations not supported by Core ML automatically run on CPU. In tests on iPhone 12, we got 35% speedup over CPU on ResNet-50. CoreML EP is convenient for fast cross-platform deployment, but for maximum performance, consider native Core ML.

When is ONNX Runtime Better Than Native Formats?

Use ONNX Runtime for prototyping, cross-platform projects, models with custom ops that coremltools cannot convert, and frequent model updates without rebuilding the conversion pipeline. If you need maximum performance on a single platform, choose the native format: on iOS — Core ML with full ANE acceleration (usually 20–40% faster than ORT+CoreML EP), on Android — TFLite + GPU Delegate (sometimes faster than ORT+NNAPI). For single-platform deployment and critical performance, native is preferable.

How to Integrate ONNX Runtime on Android and iOS

Android: Setup and Inference

// build.gradle
implementation("com.microsoft.onnxruntime:onnxruntime-android:1.18.0")

// Create session
val sessionOptions = OrtSession.SessionOptions().apply {
    // NNAPI Execution Provider for Android NPU/DSP
    addNnapi(NNAPIFlags.USE_FP16)  // FP16 mode in NNAPI
    // Or: addXnnpack(mapOf()) for XNNPACK (CPU SIMD)
    setOptimizationLevel(OrtSession.SessionOptions.OptLevel.ALL_OPT)
    setIntraOpNumThreads(4)
}

val env = OrtEnvironment.getEnvironment()
val session = env.createSession(
    context.assets.open("model_simplified.onnx").readBytes(),
    sessionOptions
)

// Inference
val inputTensor = OnnxTensor.createTensor(
    env,
    FloatBuffer.wrap(preprocessedArray),
    longArrayOf(1, 3, 224, 224)
)

val results = session.run(mapOf("input" to inputTensor))
val outputArray = (results["output"]?.value as Array<FloatArray>)[0]

// Resource cleanup — mandatory
inputTensor.close()
results.close()

Leaks from unclosed OnnxTensor and OrtSession.Result are common. In Kotlin, use use {} block: results.use { ... }.

iOS: ObjC/Swift Integration

// Package.swift or Podfile: pod 'onnxruntime-objc'
import onnxruntime_objc

// Setup
let env = try ORTEnv(loggingLevel: ORTLoggingLevel.warning)
let options = try ORTSessionOptions()
try options.setIntraOpNumThreads(4)
// On iOS — CoreML Execution Provider
try options.appendCoreMLExecutionProvider(withFlags: [.enableOnSubgraphs])

let session = try ORTSession(
    env: env,
    modelPath: Bundle.main.path(forResource: "model_simplified", ofType: "onnx")!,
    sessionOptions: options
)

// Prepare input
let inputShape: [NSNumber] = [1, 3, 224, 224]
let inputData = Data(bytes: preprocessedFloats, count: preprocessedFloats.count * MemoryLayout<Float>.size)
let inputTensor = try ORTValue(
    tensorData: NSMutableData(data: inputData),
    elementType: .float,
    shape: inputShape
)

let outputs = try session.run(
    withInputs: ["input": inputTensor],
    outputNames: ["output"],
    runOptions: nil
)

let outputTensor = outputs["output"]!
let outputData = try outputTensor.tensorData() as Data
let floats = outputData.withUnsafeBytes { Array($0.bindMemory(to: Float.self)) }

Why Quantization is Critical for Mobile Inference

Quantization reduces model size by 4× (50 MB → 12 MB), lowers memory consumption, and speeds up CPU inference by 30-40%. Dynamic quantization does not require calibration data but yields slightly less speed gain than static. In practice, we use static quantization with a representative dataset — it gives stable improvement without significant accuracy loss (0.5-2%).

What If an Operation is Not Supported?

# Check which ops NNAPI Execution Provider supports
python -m onnxruntime.tools.check_nnapi_supported_ops --model model.onnx

# If an op is not supported — it runs on CPU (fallback)
# This is not a crash but can nullify all NNAPI acceleration

To identify bottlenecks, use the ORT Profiling API. It records per-operator timing. Enable via options.enableProfiling("ort_profile") — generates JSON viewable in Chrome chrome://tracing. Profiling on target devices helps choose the optimal execution provider. For example, on one project we switched from NNAPI to XNNPACK for a model with 80% unsupported ops, reducing inference from 300 ms to 120 ms.

What Our Work Includes

• Export and simplify ONNX graph, quantize to Int8.
• Integrate ONNX Runtime on iOS and Android with optimal Execution Providers.
• Profile performance on a fleet of 10+ real devices, including older ones.
• Compare with native formats (Core ML, TFLite) and recommend the best solution.
• Documentation for building and updating the model, integration source code.
• Guarantee stable operation and lock in inference time.

Our Experience in Mobile ML

Over 5 years deploying on-device ML in commercial applications — from retail to healthcare. Completed 20+ projects with ONNX Runtime, Core ML, and TFLite. Our engineers hold Apple and Google certifications. We guarantee the model will work on all stated devices. Get a consultation on ONNX Runtime integration — we'll assess your project and propose the optimal turnkey solution. Order ONNX Runtime integration for your app — let's discuss the details.

Timeline Estimates

Basic cross-platform ONNX Runtime integration: 2–3 weeks. With EP optimization, profiling, testing on a device fleet: 4–6 weeks. Cost calculated individually after analyzing the model and performance requirements.

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.