32GB RAM is the sweet spot for local LLMs.

Yes, 32GB is enough to run models up to about 7B parameters comfortably with 4-bit quantization, and you can squeeze a 13B model if you’re careful. But ‘run’ doesn’t mean ‘run fast’ — you’ll be on CPU inference unless you have a beefy GPU, so expect slow response times.

The key constraint is total memory for the model plus context. A 7B parameter model in 4-bit needs roughly 4-5GB of RAM. A 13B needs about 8-9GB. Leave room for your OS, browser, and context window (maybe another 8-16GB). 32GB lets you load a 13B model with a decent context and still do other stuff. At 16GB you’d be fighting for space.

If you’re buying a machine specifically for local LLMs, 32GB is the minimum I’d recommend. You can run smaller models easily, and you have headroom for 13B models. For anything bigger (30B+), you’ll want 64GB or a GPU with lots of VRAM.

Final thought: Don’t expect blazing speed, but 32GB makes local LLMs practical for experimentation and casual use.