[gernb]

Is there any way to get access to a GPT-3 like API that can be run locally (color me ignorant, I know generating net for GPT-3 is huge but I have no idea how small the usable result can be stored so that usage can happen locally instead of to some cloud server

[sbierwagen]

Public models like GPT-NeoX-20B need a minimum of 45GB of VRAM. That's two 3090s, (Maybe four, five grand, depending on how much effort you spend on bid sniping ebay auctions) or a single A100 80GB. ($20,000+)

Also note that NeoX-20B is pretty good, but it's not GPT-3 quality.