Have you benched this for coding tasks, with a fallback to a larger local model, for example Qwen-3.6-27B?

Or using it for sub-tasks, where a framework with a larger primary model dispatches simpler jobs ("summarize ...", etc.) to it?