the $1,999 Mac mini runs a 30B model in near silence on 40 watts. here's the catch nobody puts in the headline

The short answer

The M4 Pro Mac mini with 48GB unified memory (~$1,999) is the best-value small local-AI machine Apple makes: it runs 30B-class models at ~40 tok/s in real use, silently, on ~40W. Its catches are…

48GB is the tier that matters. It runs 30B-class models comfortably; the base 24GB config is fine for 8–22B. Memory is soldered — buy the capacity you'll want in three years, because you can't add…

M4 Pro bandwidth is 273 GB/s — good for its class, but roughly half a Mac Studio M4 Max's and a fraction of a discrete GPU's, which shows up in prompt processing.

Power and silence are the headline wins: ~40W under load versus 350–450W for a GPU tower, near-zero noise, and tiny electricity cost.

The weakness is prompt processing (reading long inputs) — Apple silicon trails NVIDIA here, so RAG and long-document work feel slower than the token-generation numbers suggest.

Buy it for quiet, efficient local inference of mid-size models; skip it for heavy fine-tuning or long-context throughput, where a GPU or Mac Studio fits better.