Qwen2.5-Coder is the latest series of code-specific Qwen large language models, available in six mainstream model sizes ranging from 0.5 to 32 billion parameters. This instruction-tuned 3 billion parameter model brings significant improvements in code generation, code reasoning, and code fixing, built on the strong foundation of Qwen2.5 with 5.5 trillion training tokens including source code, text-code grounding, and synthetic data. The model maintains a comprehensive foundation for real-world applications such as code agents while preserving strengths in mathematics and general competencies. The model is a causal language model with 36 layers, 16 query attention heads and 2 key-value attention heads using grouped query attention, and supports a full context length of 32,768 tokens. It has been trained in both pretraining and post-training stages and uses a transformer architecture with RoPE positional embeddings, SwiGLU activation, and RMSNorm normalization.