real deal
I've been using this to code stuff that I don't know much about. all day long.
Every response always with complete code, at a whopping 7 tokens/per second, have been.. error free!
I did not experience any language shifting, only odd behavior was he would reason with himself sometimes longer then the code was..
I found early results by sharing with him my ideas or half code, let him talk to himself and then I ask for a summary of ideas to enhance... Refresh/reload, paste back his summary and make sure you don't limit his response.. go have a snack and see for yourself. I haven't looked at any reviews but I know I can't be the only one.
[tl;dr: old qwen code ..blah.. new qwen preview.. #1]
what hardware are you using to run it?
4090rtx...and 64gb system.. context length 15,000 and context batch 32,it uses 22gb video and 15gb system with ~95% gpu and 25% cpu power usage. seems to be more consistent with the low batch size. it responds back with 350 lines of error free code surprisingly. I always reload it after something like that. LM Studio has been stable to.