1、DeepSeek-Coder:When the Large Language Model MeetsProgramming-The Rise of Code IntelligenceDaya Guo*1,Qihao Zhu1,2,Dejian Yang1,Zhenda Xie1,Kai Dong1,Wentao Zhang1Guanting Chen1,Xiao Bi1,Y.Wu1,Y.K.Li1,Fuli Luo1,Yingfei Xiong2,Wenfeng Liang11DeepSeek-AI2Key Lab of HCST(PKU),MOE;SCS,Peking Universityz

2、huqh,https:/ rapid development of large language models has revolutionized code intelligence insoftware development.However,the predominance of closed-source models has restrictedextensive research and development.To address this,we introduce the DeepSeek-Coder series,a range of open-source code mod

3、els with sizes from 1.3B to 33B,trained from scratch on 2trillion tokens.These models are pre-trained on a high-quality project-level code corpus andemploy a fill-in-the-blank task with a 16K window to enhance code generation and infilling.Our extensive evaluations demonstrate that DeepSeek-Coder no

4、t only achieves state-of-the-artperformance among open-source code models across multiple benchmarks but also surpassesexisting closed-source models like Codex and GPT-3.5.Furthermore,DeepSeek-Coder modelsare under a permissive license that allows for both research and unrestricted commercial use.Fi

5、gure 1|The Performance of DeepSeek-Coder*Core contributors,ordered alphabetically by the name.arXiv:2401.14196v2 cs.SE 26 Jan 20241.IntroductionThe field of software development has been significantly transformed by the swift advancementof large language models(OpenAI,2023;Touvron et al.,2023),which

6、 have brought abouta new era of code intelligence.These models have the potential to automate and streamlinemany aspects of coding,from bug detection to code generation,thereby enhancing productivityand reducing the likelihood of human error.However,a major challenge in this field is theperformance