This covers running a (LLM) directly on a Mac Mini with 48GB of RAM. The key point is that the Mac Mini's design lets a sizable model be loaded and run for inference without a separate . 48GB is noted as enough capacity to run a mid-sized after quantization, a compression technique that shrinks a model's size.
Key points
An account of running a local LLM on a 48GB RAM Mac Mini
The Mac Mini's lets it run model inference without a discrete GPU
48GB is described as enough to run a mid-sized model after quantization