Qwen2.5-Coder Series: Open-Sourcing Powerful, Diverse, and Practical Code Models

Introduction

Today, the Qwen team is thrilled to announce the open-source release of the "Powerful," "Diverse," and "Practical" Qwen2.5-Coder series. This suite of models is designed to advance the development of Open CodeLLMs (Code Large Language Models) by providing robust coding capabilities, a wide range of model sizes, and practical applications.

Powerful: SOTA Performance in Code Generation

• Qwen2.5-Coder-32B-Instruct stands out as the flagship model, achieving state-of-the-art (SOTA) performance among open-source code models. It matches the coding capabilities of GPT-4o on several popular benchmarks:

  • EvalPlus: A benchmark that evaluates a model's ability to generate and execute code.
  • LiveCodeBench: Focuses on live coding tasks, assessing real-time problem-solving skills.
  • BigCodeBench: Tests the model's proficiency in handling large-scale codebases.

• Code Repair: Code repair is crucial for efficient programming. Qwen2.5-Coder-32B-Instruct excels in this area, scoring 73.7 on the Aider benchmark, which is comparable to GPT-4o.

• Code Reasoning: This refers to a model's ability to understand code execution and predict inputs and outputs accurately. The Qwen2.5-Coder-7B-Instruct has already shown impressive performance in this domain, and the 32B model further enhances these capabilities.

• Multiple Programming Languages: An intelligent programming assistant should be proficient in various languages. Qwen2.5-Coder is designed to handle a wide range of programming languages, making it versatile for different coding tasks.

Diverse: Range of Model Sizes

The Qwen2.5-Coder series now includes six mainstream model sizes:

• 0.5B: Ideal for resource-constrained environments.
• 3B: Balances performance and resource usage, suitable for a broad range of applications.
• 14B: Offers high performance while being more manageable than the largest models.
• 32B: The flagship model with SOTA performance.

These sizes cater to different developer needs, from lightweight applications to high-performance tasks. Whether you're working on a small embedded system or a large-scale enterprise application, there's a Qwen2.5-Coder model that fits your requirements.

Practical: Real-World Applications

Qwen2.5-Coder is not just about benchmarks; it's designed for practical use in real-world scenarios:

• Code Assistants: Integrate Qwen2.5-Coder into your development environment to provide code suggestions, error fixes, and documentation.
• Artifacts: Use the model to generate code artifacts, such as unit tests, API specifications, and configuration files.

The team has provided several examples showcasing how Qwen2.5-Coder can be applied in practical settings, making it a valuable tool for developers looking to enhance their productivity and code quality.

Getting Started

Qwen2.5-Coder is available on multiple platforms:

• GitHub: QwenLM/Qwen2.5-Coder
• Hugging Face: Qwen/qwen25-coder
• ModelScope: qwen/Qwen2.5-Coder
• Kaggle: Qwen-LM/Qwen2.5-Coder
• Demo: Try Qwen2.5-Coder in action at the Hugging Face Space

Join the community on Discord to stay updated and engage with other developers: Qwen Discord

Conclusion

The Qwen2