Frankly, right now I am way more satisfied with qwen-3-coder-420 using Cerebras inference than with those more powerful models.

Inference speed and fast feedback matter a lot more than perfect generation to me.