alibaba/MNN ? reverse-engineered prompt

Build me a lightweight AI inference engine in C++ that can run models locally on phones, PCs, and embedded devices without needing a big runtime. I want it to be fast, easy to ship, and able to handle common deep learning models, especially chat style language models and image generation models, with good support for mobile hardware acceleration when available.

Please include a simple way to load a model, run inference, and test it with a few demo apps so I can see it working end to end. It should also have clear build steps, docs, and examples for Android and iOS if possible, since I want this to feel like a real production ready toolkit rather than just a library. If you need to check current setup details, look up the latest docs online.