That's quite interesting. I hadn't thought of sparsity in the weights as a way to compress models, although this is an obvious opportunity in retrospect! I started doing some digging and found https://github.com/SqueezeAILab/SqueezeLLM, although I'm sure there's newer work on this idea.