Qwen-2 optimized for Dutch/English: continued pretraining on bilingual data, fine-tuned for conversations.