From your experience, what is one proven pattern for running quantized transformer inference on-device in Android or iOS apps using NNAPI or Core ML amid the CES 2026 edge AI buzz? In one concrete example, what latency or battery improvement did you achieve and with which tool or setting?

Running quantized transformer inference on mobile devices, whether Android or iOS, greatly enhances performance in real-time applications like image recognition and natural language processing. By leveraging NNAPI for Android or Core ML for iOS, latency and battery usage improve significantly. Model quantization, which shifts weights and activations from 32-bit floating-point to 8-bit integers, accelerates inference speed and reduces memory use, benefiting overall device efficiency.