Offline AI Assistant on iOS and Android with Llama.cpp

Imagine: an AI assistant in a mobile app works offline, all data stays on the device. No server transfers, no network latency. Cloud LLMs require constant connectivity, transmit sensitive data, and introduce delays. For medicine or finance, this is unacceptable. An on-device solution solves these issues but requires careful platform-specific integration. We integrate Llama.cpp — an LLM inference library on CPU/GPU — into iOS and Android apps. Let's dive into the technical details: from model selection to thermal management.

How to Choose a Model for an Offline Assistant?

Llama.cpp works with models in GGUF format. Popular options for mobile:

On iPhone SE 2nd gen (3 GB RAM), Llama-3.2-3B Q4 runs at the limit — OOM is possible with long contexts. A safe choice for a wide range of devices is models up to 1.5–2 GB. In one project for a financial app, we chose Llama-3.2-1B Q4_K_M, which kept memory usage under 1 GB on iPhone SE. Generation speed was 25–30 t/s, sufficient for answering queries. Thermal throttling was minimized by limiting the context to 1024 tokens.

Problems and Solutions for On-Device LLM

How to Build llama.cpp for iOS?

# Clone repository
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp

# Build via CMake for iOS
cmake -B build-ios \
    -DCMAKE_TOOLCHAIN_FILE=ios.toolchain.cmake \
    -DPLATFORM=OS64 \  # arm64 only
    -DLLAMA_METAL=ON \  # Metal GPU acceleration
    -DLLAMA_STATIC=ON
cmake --build build-ios --config Release

Result — libllama.a static library. Create a Swift Package with C-bridging header:

// llama_bridge.h
#include "llama.h"
// Wrappers for Swift-friendly API
void* llama_create_context(const char* model_path, int n_ctx, int n_gpu_layers);
const char* llama_generate_token(void* ctx, const char* prompt);
void llama_free_context(void* ctx);

n_gpu_layers — number of layers offloaded to Metal GPU. A value of -1 means all layers on GPU. On iPhone 14 with 6 GB unified memory — use -1. On devices with 3 GB — experiment: too many layers on GPU cause OOM.

Swift Wrapper for Token Streaming

import Foundation

actor LlamaSession {
    private var context: OpaquePointer?
    private var model: OpaquePointer?

    func load(modelPath: String, contextSize: Int32 = 2048, gpuLayers: Int32 = -1) throws {
        var params = llama_model_default_params()
        params.n_gpu_layers = gpuLayers

        model = llama_load_model_from_file(modelPath, params)
        guard model != nil else { throw LlamaError.modelLoadFailed }

        var ctxParams = llama_context_default_params()
        ctxParams.n_ctx = UInt32(contextSize)
        ctxParams.n_batch = 512

        context = llama_new_context_with_model(model, ctxParams)
    }

    func generate(prompt: String) -> AsyncThrowingStream<String, Error> {
        AsyncThrowingStream { continuation in
            Task.detached(priority: .userInitiated) {
                // Tokenization
                var tokens = [llama_token](repeating: 0, count: 4096)
                let nTokens = llama_tokenize(self.model, prompt, Int32(prompt.utf8.count),
                                              &tokens, 4096, true, false)

                // Inference — one token at a time
                for i in 0..<nTokens {
                    llama_batch_add(&batch, tokens[Int(i)], llama_pos(i), [0], false)
                }

                while true {
                    llama_decode(self.context, batch)
                    let nextToken = llama_sample_token_greedy(self.context, &candidates)

                    if nextToken == llama_token_eos(self.model) { break }

                    // Convert token to string
                    var buf = [Int8](repeating: 0, count: 64)
                    llama_token_to_piece(self.model, nextToken, &buf, 64, 0, true)
                    let piece = String(cString: buf)

                    continuation.yield(piece)
                }
                continuation.finish()
            }
        }
    }
}

Streaming tokens via AsyncThrowingStream — users see text as it's generated, without waiting for the full response. This is critical for UX: 10 tokens per second feels acceptable when text appears gradually.

Why Thermal Constraints Are Critical?

Llama.cpp on iPhone heats the device during prolonged generation. iOS throttling: when overheating, the system reduces clock speed, and generation speed drops from 25 t/s to 8–10 t/s. This is not a bug — it's system behavior.

Practical solution: limit the maximum context (n_ctx) to 1024–2048 for short sessions. Pause between requests. Monitor ProcessInfo.processInfo.thermalState on iOS:

NotificationCenter.default.addObserver(forName: ProcessInfo.thermalStateDidChangeNotification, ...) { _ in
    let state = ProcessInfo.processInfo.thermalState
    if state == .critical || state == .serious {
        // Pause generation, notify user
    }
}

Typical Integration Mistakes

• Context too large — choose n_ctx ≤ 2048 for mobile devices.
• Ignoring thermal throttling — monitor thermalState and pause.
• Wrong model version — verify GGUF file compatibility with your llama.cpp build.
• Missing hash verification — corrupt files cause crashes.

Android: llama.cpp via NDK

// CMakeLists.txt in jni/
add_library(llama_jni SHARED llama_jni.cpp)
target_link_libraries(llama_jni llama ggml)

// Kotlin side
class LlamaEngine {
    init { System.loadLibrary("llama_jni") }

    external fun loadModel(modelPath: String, nGpuLayers: Int): Long  // returns handle
    external fun generateNext(handle: Long, tokens: IntArray): String
    external fun freeModel(handle: Long)
}

On Android — Vulkan backend instead of Metal: include LLAMA_VULKAN=ON in CMakeLists. Supported on devices with Vulkan 1.1+, practically all with Android 10+.

Problem on Android: the process does not have a memory limit as a pooled resource — the system may kill the app (SIGKILL) when RAM is insufficient without warning. ComponentCallbacks2.onTrimMemory(TRIM_MEMORY_RUNNING_CRITICAL) — the last chance to free context before process termination.

Model Download: Progress and Verification

GGUF files weigh 1–4 GB. Download via URLSession (iOS) or WorkManager with DownloadManager (Android). SHA256 verification is mandatory: after download, compute the hash and compare with the expected one from the repository on HuggingFace. A corrupted GGUF causes a crash during header parsing or later during inference — better to catch it at verification.

The mobile neural network operates faster without network latency, which is especially important for time-critical applications. Cost savings: a fully offline solution eliminates server infrastructure expenses.

What's Included in Integration

1. Analysis of device fleet and selection of optimal quantization model
2. Building llama.cpp for iOS (Metal) and/or Android (Vulkan)
3. Development of Swift/Kotlin wrapper with asynchronous token streaming
4. Implementation of model download with progress and SHA256 verification
5. Chat UI with thermal state indication
6. Stress testing on real devices and fine-tuning of context parameters
7. Integration documentation and support during launch

Estimated Timelines

Single platform, basic chat interface with chosen model — from 3 weeks. Both platforms, multiple model choices, background download, context management — from 7 weeks. Cost is calculated individually.

Our experience: 5 years in mobile development and over 20 projects with on-device ML. We guarantee the solution works on target devices after testing. Get a consultation on model selection and project evaluation. Order integration and see the benefits of an offline AI assistant.

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.