QWEN 3.6 27B can run fine on a 16GB video card and if you give it more time it’ll be as ‘smart’ as bigger models.
As much as I'd like this to be true (don't believe all the benchmarks), in reality, using e.g. gpt 5.5 is still a lot less pain in the ass, mostly has to do with more reprompting (gpt is just smarter, oneshots stuff more often) + a lot slower (on an RTX 3090 for reference).
I've tried using it for some time, but I think I'm faster writing (better, although that's also true for gpt-5.5) code by hand, than using this (+ I need the valuable VRAM for other stuff, as I'm a graphics/shader programmer most of the time).
That said, it's already fairly impressive how much progress these smaller models have made the last year, it's usable, you can "vibe-code" at least simple stuff.