32GB is the bare minimum for local LLMs.

Yes. If you want to run anything larger than the tiny 7B parameter models, 32GB is where things actually start working.

The 7B models (like Llama 2 7B) can technically run on 16GB, but it’s tight. You’ll have to use quantized versions (lower precision), close everything else, and even then you might hit swap. 13B models? Forget it on 16GB — you’re in swap city, and at that point your LLM is slower than reading a PDF.

With 32GB you can run 7B models comfortably at 4-bit quantization, and even 13B models will fit with room to spare for your OS and browser. Higher parameter models like 30B or 70B still need more (like 64GB+), but 32GB unlocks the sweet spot of practical local inference without constant frustration.

If you build a machine today for local LLMs, 32GB is the floor. I would not build anything less.