Qwen2 marks the evolution from Qwen1.5, introducing pretrained and instruction-tuned models in five sizes, enhanced multilingual capabilities, state-of-the-art performance, and extended context lengths. All models now feature Group Query Attention and improved embedding strategies. Qwen2-72B demonstrates superior performance in various domains, surpassing Qwen1.5-110B with fewer parameters. The models excel in coding, mathematics, and handling long contexts. Significant efforts address multilingual competence, code-switching, and model safety. Openness is key, with models released on Hugging Face and ModelScope. The aim is to continue advancing Qwen2 through larger models and multimodal capabilities.
https://qwenlm.github.io/blog/qwen2/