SSD token caching (like oMLX does) goes a long way for prefill. Typical miss rates are under 5% in long coding sessions, which is mostly the output of a turn's tool use.

Prefill isnt a huge problem with an MoE, but ya, dense models are a non starter before the M5. You are still better off using DeepSeek than a local MoE, but locally running something gives you possibilities if you are on an airplane without internet, or if you want to jump guard rails, or if you want to run a home device server, etc…