Memory decides
Parameters and size
Models are described by their number of parameters: 4B, 8B, 14B, 30B, 70B (billions). More parameters usually means more capable and more memory needed. Some newer models are 'mixture of experts': a 30B model might use only about 3B parameters for each word, which makes it much faster than its size suggests, though it still needs memory for the whole file.
Quantisation
Model files are usually compressed ('quantised') to 4 bits per parameter instead of 16, cutting size by about three-quarters with a modest loss of quality. You'll see labels such as Q4_K_M or 4-bit. As a rough guide, a 4-bit model needs a little over half a gigabyte per billion parameters, plus room for your conversation.
A rough sizing guide
8 GB of memory: small models around 3–4B, fine for rewording and simple questions. 16 GB: 7–14B models, good everyday quality. 32 GB or more, or a graphics card with 16 GB+: around 30B models, which handle longer and harder tasks. Very large models (70B and up) need high-end machines. All the main apps estimate whether a model fits before you download.
Macs, PCs and graphics cards
Apple Silicon Macs share one pool of memory between the processor and graphics, so a 32 GB MacBook can run models a 32 GB Windows laptop can't run quickly. On Windows and Linux, an NVIDIA graphics card with plenty of video memory (VRAM) makes the biggest difference; AMD cards are increasingly supported. Without a graphics card, small models still run on the processor, just slowly.
Do you know how much RAM your computer has? Where would you look to find out?