Abstract

This report introduces the Qwen2 series, the latest addition to our largelanguage models and large multimodal models. We release a comprehensive suiteof foundational and instruction-tuned language models, encompassing a parameterrange from 0.5 to 72 billion, featuring dense models and a Mixture-of-Expertsmodel. Qwen2 surpasses most prior open-weight models, including its predecessorQwen1.5, and exhibits competitive performance relative to proprietary modelsacross diverse benchmarks on language understanding, generation, multilingualproficiency, coding, mathematics, and reasoning. The flagship model, Qwen2-72B, showcases remarkable performance: 84.2 onMMLU, 37.9 on GPQA, 64.6 on HumanEval, 89.5 on GSM8K, and 82.4 on BBH as a baselanguage model. The instruction-tuned variant, Qwen2-72B-Instruct, attains 9.1on MT-Bench, 48.1 on Arena-Hard, and 35.7 on LiveCodeBench. Moreover, Qwen2demonstrates robust multilingual capabilities, proficient in approximately 30languages, spanning English, Chinese, Spanish, French, German, Arabic, Russian,Korean, Japanese, Thai, Vietnamese, and more, underscoring its versatility andglobal reach. To foster community innovation and accessibility, we have made the Qwen2model weights openly available on Hugging Face1 and ModelScope2, and thesupplementary materials including example code on GitHub3. These platforms alsoinclude resources for quantization, fine-tuning, and deployment, facilitating awide range of applications and research endeavors.