if you're having some cards already try to explore what people are doing with weight streaming and model compression, I got good results on older cards with far better models than the 16B

and the new thing will be "will it run qwen" :D