We must have very different workflows, I am curious about yours. What tools are you using and how are you guiding Qwen3-Coder? When I am using Claude Code, it often works for 10+ minutes at a time, so I am not aware of inference speed.
I‘ve found that the best way for myself to do LLM assisted coding at this point in time is in a somewhat tight feedback loop. I find myself wanting to refine the code and architectural approaches a fair amount as I see them coming in and latency matters a lot to me here.
Indirectly, it sounds like you're aware about the inference speed? Imagine if it took 2 minutes instead of 10 minutes, that's what the parent means.