DeepSeek-V2-Chat-0628 is the improved version of DeepSeek-V2, specifically designed for dialogue generation tasks. It has outstanding performance on the LMSYS Chatbot Arena Leaderboard, ranking 11th overall, particularly in programming tasks and challenging prompts. The model has significant improvements in various evaluation metrics, such as HumanEval, MATH, BBH, IFEval, and Arena-Hard, and has optimized instruction following capabilities in the "system" field, greatly improving user experience.