Since I'm just a pleb with a 5090, I run GPT-OSS 20B a lot, since it fits comfortably in VRAM with max context size. I find it quite decent for a lot of things, especially after I set reasoning effort to high and disabled top-k and top-p and set min-p to something like 0.05.

For the Qwen3-VL, I recently read that someone got significantly better results by using F16 or even F32 versions of the vision model part, while using a Q4 or similar for the text model part. In llama.cpp you can specify these separately[1]. Since the vision model part is usually quite small in comparison, this isn't as rough as it sounds. Haven't had a chance to test that yet though.

[1]: https://github.com/ggml-org/llama.cpp/blob/master/tools/serv... (using --mmproj AFAIK)