Integrate Core ML on iOS: Offline AI Without the Cloud

You train an image classification model on PyTorch—accuracy 95%, but how do you run it on iOS without latency and without sending data to the cloud? Core ML with on-device inference solves this. We integrate Core ML models into iOS apps for fully offline AI. Inference speed—single-digit milliseconds, data stays on the device, no network latency. Our team—7 years in mobile development, over 50 successful Core ML integrations. Savings on server inference can reach 60% (over 200,000 ₽ per year for an average project). We guarantee integration quality, confirmed by Apple Developer certifications.

How conversion works: from weights to .mlpackage

Most modern models arrive as PyTorch checkpoints or ONNX files. We convert via coremltools—Apple's official Python package:

import coremltools as ct
import torch

# Suppose we have a PyTorch image classification model
model = MyModel()
model.load_state_dict(torch.load("model.pth"))
model.eval()

# Tracing—need to pass example input
example_input = torch.zeros(1, 3, 224, 224)
traced = torch.jit.trace(model, example_input)

# Conversion
mlmodel = ct.convert(
    traced,
    inputs=[ct.ImageType(
        name="input_image",
        shape=(1, 3, 224, 224),
        color_layout=ct.colorlayout.RGB,
        bias=[-0.485/0.229, -0.456/0.224, -0.406/0.225],  # ImageNet normalization
        scale=1/(255.0 * 0.229)  # built into model, no need to do in Swift
    )],
    outputs=[ct.TensorType(name="class_probabilities")],
    compute_precision=ct.precision.FLOAT16,  # for ANE
    minimum_deployment_target=ct.target.iOS16
)

mlmodel.save("MyClassifier.mlpackage")

FLOAT16 + minimum_deployment_target=iOS16 activates the Apple Neural Engine. On iPhone 14, this is 4–8× faster than GPU for inference, with significantly lower battery consumption. According to Apple's Core ML documentation, ANE accelerates inference 4–8× compared to GPU. On older iOS versions, the same model runs via Metal GPU.

How to convert a model from PyTorch to Core ML?

Dynamic shapes—models with torch.Size([batch, seq_len, hidden]) where seq_len is not fixed break torch.jit.trace. Solution: ct.RangeDim for variable sizes or define multiple configurations via ct.EnumeratedShapes.

# Variable sequence length
flexible_shape = ct.Shape(shape=(1, ct.RangeDim(1, 512), 768))
mlmodel = ct.convert(model, inputs=[ct.TensorType(shape=flexible_shape)])

Unsupported operations—for example, custom CUDA kernels. coremltools throws NotImplementedError. Path: either rewrite the operation using standard PyTorch primitives, or add a custom layer via C++/Swift extension.

Error Unsupported model format when loading .mlpackage on x86 simulator—the simulator uses CPU fallback, some FLOAT16 operations are not supported. Test accuracy only on a real device.

Loading and running on iOS

import CoreML
import Vision

// Load model (once at startup)
let config = MLModelConfiguration()
config.computeUnits = .all  // ANE + GPU + CPU

// .mlpackage loaded from bundle
guard let modelURL = Bundle.main.url(forResource: "MyClassifier", withExtension: "mlpackage"),
      let model = try? MyClassifier(contentsOf: modelURL, configuration: config) else {
    fatalError("Failed to load model")
}

// Inference—on background thread
DispatchQueue.global(qos: .userInitiated).async {
    do {
        let input = MyClassifierInput(input_image: cgImage)
        let output = try model.prediction(input: input)
        let probs = output.class_probabilities
        // probs — MLMultiArray, get value: probs[0].doubleValue
    } catch {
        print("Inference error: \(error)")
    }
}

Model loading takes ~100–300 ms (depends on size). Do not load it in viewDidLoad—load once at app startup or first use, keep in memory while needed.

Why on-device ML is faster and safer than cloud?

On-device ML eliminates network latency, preserves user data privacy, and works offline. You don't pay for server inference and don't depend on internet connection. For tasks where response speed is critical (e.g., real-time video processing), device is the only sensible option.

Performance on real devices:

For profiling, use Xcode Instruments → Core ML Instrument.

Vision Framework as a wrapper

For computer vision tasks, VNCoreMLRequest is more convenient—Vision handles input resizing, image orientation, coordinate transformations:

let coreMLModel = try VNCoreMLModel(for: model.model)  // .model — MLModel from generated class

let request = VNCoreMLRequest(model: coreMLModel) { request, error in
    guard let results = request.results as? [VNClassificationObservation] else { return }
    let topResult = results.sorted { $0.confidence > $1.confidence }.first
    print("\(topResult?.identifier ?? "?") — \(topResult?.confidence ?? 0)")
}
request.imageCropAndScaleOption = .centerCrop  // or .scaleFit

let handler = VNImageRequestHandler(cgImage: inputCGImage, options: [:])
try handler.perform([request])

VNCoreMLRequest automatically solves the input size mismatch problem—you pass an arbitrary image, Vision resizes it to the model's expected size. Without Vision, you'd have to do this manually via vImage or CIImage.

What's included in the work

• Documentation on conversion and integration, including description of all steps and tools used.
• Access to the repository with conversion source code and integration examples.
• Training of the client's team on working with Core ML, profiling, and model updates.
• Support for one month after integration to resolve any issues.

How to update a model without updating the app?

Core ML supports loading a model from an arbitrary URL, not only from bundle. This allows updating the model via server:

// Load mlpackage from documents directory
let documentsURL = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]
let downloadedModelURL = documentsURL.appendingPathComponent("updated_model.mlpackage")

if FileManager.default.fileExists(atPath: downloadedModelURL.path) {
    let model = try MyClassifier(contentsOf: downloadedModelURL, configuration: config)
} else {
    // Fallback to bundle
}

Download model over network via URLSession, save to Documents, verify via SHA-256 hash before use.

Our approach: from analysis to deployment

• Analysis of the original model (framework, weights, structure).
• Conversion to Core ML with optimization of precision and compute units.
• Optimization for target devices (profiling on real devices).
• Integration into the app: loading, caching, fallback on errors.
• Setting up remote model update (optional).
• Documentation and team training.

Process stages:

1. Analytics—receive weights, assess complexity, choose conversion strategy.
2. Conversion—create .mlpackage, resolve operation and dimension issues.
3. Profiling—measure speed and energy consumption on several iPhone generations.
4. Integration—embed into SwiftUI/UIKit, add error handling.
5. Deployment—publish via App Store, set up remote update.

Estimated timelines

Conversion of an existing model + basic iOS integration—1–2 weeks. Complex model with non-standard operations, multiple inputs/outputs, remote update—3–5 weeks. Cost is calculated individually for each project. We'll assess the task in one business day—just send the weights and a description of the task.

Contact us for a consultation on Core ML integration into your project. Request a preliminary analysis of your model—we'll find the optimal solution.

Learn more about conversion in the coremltools documentation.

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.