I have a 16 GB card personally and want to keep my agent fully loaded on the GPU whenever possible so here is my experiences:
Unsloth Dynamic 3 UD-Q2_K_XL So far, Qwen 3.8 27B seems to be performing well, but I may have to offload to CPU or cut into context length if I want to up the quant
Unsloth Dynamic 2 UD-Q2_K_XL So I have tried Qwen 3.6 35B-A3B at this quant and it had some more arithmetic errors than I found acceptable, which led me to using:
Unsloth Dynamic 2 UD-IQ3_XXS I know it's not a huge step up but it is, in my experience, less prone to errors and better at catching errors it does make.
If you have experience with this topic, let me know your thoughts. If not, I hope you find it helpful :)