JustVugg/Colibri ? reverse-engineered prompt

Build me a tiny C app that can run huge MoE models on a normal machine, even when most of the model has to live on disk and stream in as needed.

I want a command line chat mode, a local server mode, and a simple web dashboard so I can talk to the model and watch what the engine is doing while it runs. It should treat VRAM, RAM, and storage like one memory system, keep the model behavior correct, and not silently change the model just to make it faster. If possible, make it work with CPU and GPU options on the same code path, and show useful live stats like memory use, token speed, and what experts are active.

Keep it small, fast to launch, and easy to build from source. If you need to, look up current docs online for any model or hardware details, but keep the experience simple for someone who just wants to run it on hardware they already own.