Start off with the Tinyllama model, it's under 1gb. It will even run on raspberry pi so on real PCs it rips even on CPU. You need a "quantized" model, they are distributed as GGUF files.

I would recommend 5 bit quantized. The less bits, the stupider to put it simply, and Tinyllama is already pretty stupid. But it's still impressive for what it is, and you can learn the jargon which is the hard part.

Fastest software to run the model on is llama.cpp which is a rewrite from python to C++. Use -ngl to offload layers from cpu to GPU.

Not sure what system you're using, most AI development is done on Linux so if you're on Windows I can't guarantee anything will work.

Working right now on making a voice assistant for my house that can read all my MQTT data and give status reports, it's neat when you get it running. Fun to tweak it with prompts and see what it can do. Tinyllama can't seem to reliably handle MQTT and JSON but slightly smarter models can with ease.