AI Integration into Mobile Applications
Mobile AI splits into two fundamentally different approaches: cloud inference (API request to server) and on-device inference (model runs on phone). The choice depends on latency requirements, privacy, and model size.
Cloud AI for Mobile
The simplest approach: mobile app → REST API → LLM/ML model on server → response. Suitable for complex tasks where the model doesn't fit on the device. Drawbacks: latency (100–2000 ms), network dependency, server costs.
Stack: iOS (URLSession), Android (Retrofit/OkHttp). Streaming responses for LLM (SSE/WebSocket).
On-Device AI
Model runs locally — privacy, offline operation, zero latency.
iOS / Core ML:
- Conversion via coremltools (PyTorch → Core ML)
- Neural Engine on iPhone 12+ — significant acceleration
- Create ML for training simple models directly in Xcode
Android / TensorFlow Lite:
- TFLite + NNAPI for hardware acceleration
- GPU delegate for Vision tasks
- Hexagon DSP delegate on Qualcomm
Practical on-device capabilities (2025)
| Task | Platform | Model | Performance |
|---|---|---|---|
| Image classification | iOS/Android | MobileNetV3 | <10 ms |
| Object detection | iOS/Android | YOLOv8n | 20–50 ms |
| Text classification | iOS/Android | DistilBERT quantized | 50–150 ms |
| Small LLM | iOS (Neural Engine) | Llama 3.2 3B | 15–30 token/sec |
| Speech recognition | iOS/Android | Whisper tiny | Real-time |
Development Pipeline
Weeks 1–3: Approach selection (cloud/on-device/hybrid). Inference prototype.
Weeks 4–7: Model optimization (quantization, pruning). Native integration.
Weeks 8–10: AI feature UX. Error handling. Graceful degradation.