youssofal/MTPLX ? reverse-engineered prompt

Build me a native Mac app plus a simple command line tool for Apple Silicon that lets me run local language models much faster by using the model’s own multi token prediction, not a separate draft model. I want it to feel easy for a normal Mac user, so the app should check my hardware, suggest a model that fits my memory, download it, set everything up, and tune itself to find the fastest safe settings on my machine.

I also want a local server that works with OpenAI compatible and Anthropic compatible clients, with streaming chat, tool use, health and metrics endpoints, and a way for other apps to connect to the same running model without loading it twice.

Please include a clean dashboard with live tokens per second, acceptance rate, cache status, and system pressure, plus a built in chat UI with file attachments and optional web search. If it fits, add benchmarking, model inspection, and a workflow for converting and testing your own compatible models. Look up current docs online if you need to.