I spent a day reading memory numbers straight from macOS and llama.cpp on my M4 Mac mini. What I found:
The GPU gets 11.8GB, not 16
Gemma 4 12B fits with room to spare
Qwen 27B in 4-bit doesn’t come close
One full round of model downloads was silently corrupted, and the fix is in the post
Every number comes from my own machine. The script is printed in full, and there are no affiliate links.
If this stops you from buying the wrong Mac, a coffee keeps the testing going 🙏
