Route 1: a graphics card
A desktop graphics card has very fast memory, so any model that fits in its video memory runs quickly. The video memory size is the limit: 8 GB handles models up to about 9 billion parameters, and 16, 24 and 32 GB cards fit progressively larger ones. On our own RTX 4060 (8 GB), Qwen3.5 9B writes about 45 tokens per second, well above comfortable reading speed. Our graphics card guide compares 16, 24 and 32 GB cards at 2026 prices.
Route 2: a compact unified-memory machine
Machines such as Apple's Mac Studio, NVIDIA's DGX Spark and mini desktops with AMD's Ryzen AI Max chip give the graphics processor up to 128 GB of memory or more. They are slower per gigabyte than a big graphics card, but they fit models no single consumer card can hold.
Route 3: the processor alone
A modern desktop processor with fast DDR5 memory can run small models and mixture-of-experts models at usable speeds without any graphics card. It is the cheapest way to start, and a good way to learn what you need before you spend more.