I have to say, Qwen3 (running locally) is incredibly good at photo descriptions. I'm pleasantly surprised at how detailed and sensible they are.
@FreakyFwoof Are you saying Qwen3vl is better than Qwen3.5?
@chikim No, the one you told me to use. qwen3.5:35b
@FreakyFwoof Oh ok, you said qwen3, so I thought it was 3, not 3.5. lol
@chikim It runs well, even on my now ancient M1 Max.
@FreakyFwoof @chikim By the way, 35B is a sparse model and some say the 27B works better. It could be a little faster, too.
@rojun @chikim O good to know, thanks.
@FreakyFwoof @rojun The quality from 27b might be little better but absolutely runs slower than 35b because it's a dense model. The 35B model is MoE architecture with only 3B active parameters. This is especially more important for Macs because Apple Silicons with unified memory have much slower memory bandwidth compared to NVidia cards.
@chikim @FreakyFwoof Ah, makes sense