32GB is a solid upgrade for local LLMs.

Yes. If you’re running Llama 2 7B (or similar 7B models), 32GB RAM is not overkill—it’s where things start to get comfortable.

The model itself, quantized to 4-bit, eats about 5–6GB. Then you need room for your operating system, browser, and any other apps. Most importantly, you need memory for the context window—the longer your conversation, the more RAM it chews up. 16GB can work but you’ll be swapping to disk the moment you try a decent chat length. 32GB gives you breathing room.

Keep in mind: RAM helps with loading the model and handling context, but inference speed (how fast it generates tokens) is mostly about your GPU or CPU compute. If you’re running on CPU, faster RAM helps a bit, but the real bottleneck is the processor. If you have a half-decent GPU with 8–12GB VRAM, that’s actually more impactful than system RAM.

Bottom line: If you have a solid CPU/GPU and are bumping against 16GB limits, 32GB is a great upgrade. Future-you will appreciate not closing everything just to chat with a local model.