Is it something wihch would make PS5 attractive to run local LLM models? I've heard Mac Mini shared CPU/GPU memory makes it popular for that.
No, unified memory is not a panacea. The PS5 GPU is mostly raster-focused and doesn't have good acceleration options outside Vulkan compute.
Your best ~$300-400 option for inference is still the RTX 3060 AFAIK.
The Mac studio also comes with up to 512GiB RAM.
The PS5 only has 8GiB.