Thoughts
#3
by
HAV0X1014 - opened
I like this model! Its writing style isn't terrible and doesn't make me roll my eyes every time I see a repeated line. I use it with thinking disabled (--reasoning-budget 0 in llama.cpp) and I've noticed that it ignores the first prompt the user sends sometimes. I'm not sure if that's a qwen thing, a finetune thing, or a lack of thinking thing.
I like that it can behave as a character and actually be helpful for coding. I've used some mistral 24b models that could write really well but couldn't do any assistant tasks whatsoever. The qwen 3.5 base really helps here.
I think a full-er finetune on more character + assistant tasks would make this even better, if that's how it works.