antirez/ds4 ? reverse-engineered prompt

Build me a small native Mac tool that lets me run DeepSeek V4 Flash locally on Apple Silicon using Metal. I want it to feel finished enough that I can download the supported model, run make, and then either chat in the terminal or start a local HTTP server for coding agents to talk to.

Please keep it focused on this one model only, not a general model runner. Add a model download script for the q2 version for 128 GB Macs and q4 for bigger machines, with resume support and a simple default model link. The command line app should support a one prompt mode and an interactive chat mode. The server should expose a simple API and make it clear how to start it.

Handle long context and disk based KV cache as first class features. Include help text, basic tests or validation hooks, and friendly errors if someone uses the wrong GGUF or tries to run on unsupported hardware. Look up current docs online if you need to.