32GB AI PCs: Every System Compared

Enough for the models most people actually want

Thirty-two gigabytes runs a 24B model comfortably and a 32B model at a reduced context. That matters because the 24B to 32B range is where open models became genuinely competent at reasoning and code, and it is the size most people have in mind when they picture running something locally.

This band is crowded and varied: desktop RTX 5090 towers sit here on VRAM, alongside Macs and Ryzen machines reaching the same figure through unified memory. Bandwidth across this page spans more than a factor of five, so two rows with the same memory number can differ enormously in how quickly they answer.

Read the bandwidth column

On this page in particular, sorting by memory alone will mislead you. A 32GB discrete card will generate several times faster than a 32GB unified pool on any model that fits both. The unified machine wins only when you later want something that does not fit 32GB at all.