Fine-Tuning LLMs for Mobile Apps: Turnkey

Fine-Tuning LLMs for Mobile Applications Turnkey

We are a team of mobile developers with extensive experience, having completed over 50 projects fine-tuning LLMs for iOS and Android. We guarantee a minimum 40% reduction in hallucinations, and our expertise is proven by over 50 successful projects. We've encountered situations where a base model hallucinates in 30% of responses or breaks JSON formatting. Our experience shows: fine-tuning reduces hallucinations by 3–5x and increases accuracy by 40–60%. LLM fine-tuning is our core expertise. Clients typically save 40% on API costs after fine-tuning due to fewer retries. We work with both the OpenAI Fine-Tuning API and open-source models like Llama 3, Mistral, and Gemma 2. Our model integration services ensure seamless deployment. Our services are tailored for mobile app AI needs.

When Prompt Engineering Falls Short

Three scenarios where fine-tuning becomes justified:

Format determinism. The model must return strictly structured JSON with custom fields specific to your domain. Even with few-shot examples in the prompt, the base model periodically breaks the schema or adds extraneous fields. After fine-tuning on 5,000–10,000 examples, format errors disappear almost completely.

Domain terminology. A medical app with ICD-10 terms, a legal assistant with article numbers, fintech with internal product codes—the base model gets confused or interprets abbreviations generically. Fine-tuning on a corpus of your documents resolves this.

Style and tone. Brand voice is a real business need. If your assistant must respond in a specific character style or with a certain degree of formality, it's cheaper to embed this into weights than to carry it in every request via a system prompt.

Fine-tuning is 3-5x better at reducing hallucinations than prompt engineering alone. In fact, fine-tuning improves accuracy by 40–60% compared to base models, outperforming prompt engineering.

How to Prepare a Dataset for Fine-Tuning?

80% of fine-tuning success is determined by the quality of training data. As noted in the official OpenAI documentation, the minimum volume for noticeable results is 50–100 examples; a realistic volume for production is 500–2,000 pairs. Formatting for the OpenAI Fine-Tuning API (gpt-4o-mini):

{"messages": [
  {"role": "system", "content": "You are a medical app assistant. Answer symptom questions briefly and safely."},
  {"role": "user", "content": "My resting heart rate is 45 bpm."},
  {"role": "assistant", "content": "Bradycardia. Normal for trained athletes. If accompanied by dizziness or fainting, consult a cardiologist."}
]}

When automatically generating datasets via GPT-4, manual validation is mandatory: automatically created examples reproduce base model errors. For open-source models, the dataset is prepared in Alpaca or ShareGPT format and passed to Hugging Face datasets. We handle dataset generation from scratch. Custom model training is available for unique requirements.

Choosing an Approach: OpenAI vs Open-Source

For most mobile products, OpenAI Fine-Tuning is the fastest path to results. If data cannot leave your perimeter (medical, finance), use open-source with deployment on your own server or local execution via CoreML / llama.cpp. We also offer Llama 3 fine-tuning services.

Integrating a Fine-Tuned Model into a Mobile App

After training, the model receives an ID. In the mobile app code, the only change is to substitute this ID for the base one:

let request = ChatCompletionRequest(
    model: "ft:gpt-4o-mini:org:id",
    messages: conversationHistory,
    maxTokens: 256,
    temperature: 0.3
)

data class ChatRequest(
    val model: String = "ft:gpt-4o-mini:org:id",
    val messages: List<Message>,
    val max_tokens: Int = 256,
    val temperature: Double = 0.3
)

At the API level, there is no difference—same REST endpoint, same response format.

Quality Evaluation and Iterative Improvement

Fine-tuning is not a one-time operation. Standard cycle:

1. Baseline measurement on a test set (15–20% of dataset, held out before training)
2. Training → launch A/B test in the app on 10% of traffic
3. Collect user feedback (likes/dislikes, corrections to answers)
4. Augment the dataset with problematic examples
5. Retrain

We perform A/B testing model selection to ensure best performance.

OpenAI Dashboard displays training loss and validation loss per epoch. Overfitting is visible when curves diverge—validation loss rises while training loss falls. In that case, reduce the number of epochs or increase the dataset.

Estimated Timelines and Resources

Full cycle from audit to production: 3–8 weeks. With ready annotated data, from 1 week. Our fine-tuning services start at $1,500 for a basic dataset of 500 examples, with full project costs ranging from $3,000 to $10,000.

What's Included

• Audit of current prompts and identification of bottlenecks
• Dataset preparation and labeling (500–2000 examples)
• Model training with metric monitoring
• Integration of the fine-tuned model into the mobile app
• A/B testing and dataset augmentation
• Documentation and team training
• 1 month of support

Mobile AI applications require specific optimization. Get a consultation and timeline estimate for your project. Contact us — we'll discuss the details and choose the best approach.

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.