Not enough VRAM to get a model running that can do real work. The minimum is 32gb currently to run qwen3.8:27b with about a 96k contact window. Anything smaller than that is not going to be particularly useful, in my experience.

I'm currently testing Qwen3.6-35B-A3B-UD-Q3_K_M with a RTX 3060 12GB and 32GB of RAM. Can't say is super fast but not bad in my last test it managed to analyze a project kinda big and pointed two options for how to implement what I asked. My usage is kinda the same as you said, kinda like a toy to p...

Depends on how my ddr5 he has too, many of the MoE models can offload experts. He should look at free token to do inference. But I think he probably needs to look at qwen3.5 class models, maybe 9b or 35b-a3b