i only know that the gpt-oss model, 20b weighs in about 12gb, and is about the biggest i can run on a 16gb GPU.

i can also tell you that you need twice as much, at least, to do programming level reasoning. the cloud services with gpt5 and all those, they are running on 128gb+ devices. these are coming to the domestic market end of this year. actually training them, though, it's just like bitcoin mining. the more, the sooner the reward.