As I understand it, the language model includes a training set, which is why even the "smallest" are measured in gigabytes. Qwen2.5 knows what all of those are, so you might try setting that as your model and see if you get better results with your scenarios.