DeepSeek-Coder-V2 is an open-source expert mixture model (MoE) specifically designed for code language. Its performance is comparable to GPT4-Turbo. It excels in code-specific tasks while maintaining strong performance in general language tasks. Compared to DeepSeek-Coder-33B, the V2 version demonstrates significant improvements in code-related tasks and reasoning capabilities. Furthermore, it expands its supported programming languages from 86 to 338 and increases the context length from 16K to 128K.