sqliteai/waste ? reverse-engineered prompt

Build me a small C project that can run a huge local AI model from disk without needing to load all the weights into RAM.

I want a command line tool called waste that can open a converted model file, answer a text prompt, and keep the shared model parts in memory while streaming the expert weights from fast storage as needed. It should work on macOS or Linux, build with make, and not depend on Python, CUDA, BLAS, or other runtime libraries for inference. Please include a simple test suite that can run without downloading any real model, plus a couple of examples for text and image input.

If it makes sense, also add a small library version so other programs can embed it, and a basic server mode for local use. Keep the code focused on the model loading, memory budgeting, caching, and fast disk reads, and make sure it can run with limited RAM as long as the model is on internal NVMe. Look up current docs online if you need to.