Qwen3.6-35B-A3B is pretty amazing. I'm using it with 96k context on 24GB VRAM through ollama.

Gemma 4 E4B and Qwen 3 4B are pretty good, but fine-tuning makes them really good. There are tradeoffs at this size, so you'll have to find (or make) a finetune that does what you need.

Maybe bonsai 8b would make the duo, if you do try it, pls post here as I'm a bit curious too.

granite 4