16GB is enough for Llama 3.1 8B at 4-bit. 8GB will make you suffer.

Yes, 16GB works. No, you won’t swap heavily—if you keep other apps light. 8GB is a hard no for that model.

The 8B parameter model at 4-bit quantization takes about 5-6GB just to load. Then you’ve got overhead for inference, context, and your OS. That leaves you maybe 9-10GB free on a 16GB machine. Chrome’s a pig, but you can run the model and a browser just fine.

With 8GB, you’re looking at maybe 2GB free after loading the model. That’s swap city. The model itself will thrash so hard it’s borderline unusable. You could try a smaller quant like 3-bit or a smaller model (7B/3B), but the 8B at 4-bit on 8GB is a bad time.

If you’re buying RAM, go to 32GB if you want to run larger models or multitask. But 16GB is the minimum for comfortable use with that specific model. Future You deserves smoother inference.