Strongly recommend https://github.com/Neroued/ninfer, which can pull ~180 TPS on 5090 with 3.8, and 500 (!) with 3.6 35B-A3B.

I've been running 27B a lot, I am honestly shocked at how well it performs. It's mind blowing how well the small qwen models (and particular, the 3.8 model) runs locally. It can create some really impressive toy coding projects. I haven't really done too much integrating into my actual workflows because I pay $100 for Claude Max, but I can see it being pretty helpful in that.