Qwen 1.5B Crawls on CPU

Small models are still painfully slow when the runtime falls back to CPU on weak mobile hardware.

• –On a Redmi 12-class Snapdragon 4 Gen 1 phone, memory bandwidth and thermals can dominate long before parameter count does.
• –llama.cpp officially supports Adreno OpenCL, but Snapdragon Hexagon offload is still marked in progress, so Android acceleration is backend-dependent.
• –0.12 tok/s is consistent with a CPU-only path, even for a 1.5B model, if the app is not actually offloading work.
• –For local mobile LLMs, backend support and quantization matter as much as model size.