General go to is pair of nvidia cards with 24gb of ram a piece. Should be enough to run stuff like Mixtral 8x7b in 8 bit precision which is good enough. That b

General go to is pair of nvidia cards with 24gb of ram a piece. Should be enough to run stuff like Mixtral 8x7b in 8 bit precision which is good enough. That being said, single 24gb card is fine enough 4bit precision models if you are using it for basic coding assistance. If you are interested in inference only, not training, its not really worth it to invest in cards. Use the online inference tools. And for training, even a pair of 4090s aren't going to be that good without a good CPU and lots of RAM to keep the cards fed as much as possible.