[Official FAQ] Share your experience or follow-up questions below.

For local LLMs, a good priority order is: VRAM capacity > memory bandwidth > software support > raw compute.

• VRAM first: it decides which models fit entirely on the GPU. Once a model spills onto the CPU, speed drops sharply. Very roughly, for 4-bit models: 12 GB suits 7B–8B, 16 GB can handle around 14B, and 24 GB-class cards can reach the 30B range. Bigger needs multiple GPUs or a large-memory machine.
• Bandwidth: token generation is mostly limited by memory bandwidth, so at equal VRAM a higher-bandwidth card is usually faster.
• NVIDIA is the easiest path: CUDA has the widest support and the most tutorials. AMD can work via ROCm or Vulkan, but support depends on the exact model and OS, so verify that your software supports your card before buying.
• Used high-VRAM cards: a used previous-generation 24 GB card is a popular value choice. Check condition, warranty, power draw and cooling.
• Do not forget the rest of the system: power supply, case size, cooling, and enough system RAM (at least as much as your VRAM is a sensible target).

Money-saving ideas

• Test models on hardware you already own to learn what size you really need.
• If you only need occasional heavy runs, renting a cloud GPU or using a hosted API may be cheaper.

Prices change quickly, so none are listed here. What is your budget and main use case? Tell us and we will help you shortlist.