Last I checked (around the time that LLAMA v3 was released), the performance of local models on CPU also was pretty bad for most consumer hardware (Apple Silicon excepted) compared to GPU performance, and the consumer GPU RAM situation is even worse. At least, when talking about the models that have performance anywhere near that of ChatGPT, which was mostly 70B models with a few exceptional 30B models.

My home server has a 3090, so I can use a self-hosted 4-bit (or 5-bit with reduced context) quantized 30B model. If I added another 3090 I’d be able to use a 4-bit quantized 70B model.

There’s some research that suggests that 1.58 bit (ternary) quantization has a lot of potential, and I think it’ll be critical to getting performant models on phones and laptops. At 1.58 bit per parameter, a 30B model could fit into 6 gigs of RAM, and the quality hit is allegedly negligible.