Depends on the job. 12GB is a huge limitation. You can't have the hot MoE models of the month (Qwen3.6 35B A3E and Gemma4 26B A4E) without serious compromises on quantization (quality of results), context size (type of tasks) or speed (offloading to RAM kills performances, you can't interact, you basicly enter a message and come back minutes later).

I find the stricter the contraints, the harder you have to experiment with models and configurations yourself, all depends on your needs and expectations. If your workload allows for (or even encourages) long waits, you can afford running off RAM, and you get in the realm of < 5t/s go with Qwen3.6 35B A3E or Gemma4 26B A4E. I'm experimenting on a remote PC with 32GB of unused RAM (bought when RAM was cheap). It would be sitting there doing nothing otherwise. I'm experimenting with Qwen3.6 35B A3E Q5_K. It's relatively fast, among snails, but I find it overthinks a bit too much, which is annoying when you're watching 3 or 4 t/s production. Gemma tends to think less, it's a tad slower but gets to the point faster.

If you want speed, you have to go with smaller models. I barely tested those, I've heard Gemma4 E4B is good. I've heard rumors they are cooking Qwen3.6 smaller models as well.