Running Vision-Language Models On-Device in Android

Technical deep-dive into running VLMs (LLaVA/MobileVLM-class) on Android — covering the dual-model architecture (CLIP vision encoder + language decoder), INT4/INT8 quantization trade-offs for vision towers vs language heads, CameraX frame buffer pipeline integration, GPU delegate for the vision encoder with NNAPI fallback for the LM decoder, memory pressure management under sustained dual-model inference, thermal throttling strategies, and the Kotlin coroutine streaming pipeline that returns structured responses while keeping the camera preview at 60fps