You can see that Qwen3 does worse than Qwen2.5 on our benchmark. Reason is it's never been pretrained for FIM / autocomplete.