I just got a Mac mini with an M5 Pro (base CPU config) and 48GB of unified memory. I want to run as much AI locally as makes sense, but it's hard to separate what actually works day to day from benchmark hype.

What I'd like to do:

  • Coding: autocomplete and agentic workflows with local models
  • Image generation
  • Anything else genuinely useful: transcription, RAG over my notes, etc.

Questions:

  1. Which models and quantizations fit comfortably in 48GB and are actually usable, not just "it runs"?
  2. Which runtime and tools: MLX, llama.cpp, Ollama, LM Studio, ComfyUI, something else?
  3. What's your setup, and what did you give up on and go back to cloud APIs for?

Real tokens/sec numbers and "this was a waste of time" stories are especially welcome.