so Qwen just dropped Qwen 3.8 27B and its making waves on HN right now
its an open source 27B parameter model from Alibaba with Apache 2 license. vision capable too. runs locally on a decent laptop which is wild for that size
the benchmarks apparently beat their own bigger closed model from a few months ago. 27B beating models 5x its size
the funny part is it defaults to overthinking. like burning 22k reasoning tokens to draw a pelican on a bicycle. 21 minutes of thinking. you gotta turn down the reasoning effort or it goes crazy
since i work with AI models daily at Megallm this caught my eye. a 27B open model that actually competes with closed ones is a big deal for local inference and self hosting
anyone tried running it yet? thoughts on the overthinking thing? worth setting up on a local machine?
Top comments (0)