If you’re able to run qwen3-coder, have you thought about 3.6 27B or 35B? Looking at benchmarks, 3.6 looks its gained a lot over qwen3-coder