Qwen2.5-Coder Series: Strong Performance, Diverse Models, High Practicality

The Qwen2.5-Coder series has been released as open source. This series emphasizes strong performance, diverse model sizes, and practicality as its core values.

The flagship model, Qwen2.5-Coder-32B-Instruct, achieves the best performance among open-source models on major code generation benchmarks such as EvalPlus, LiveCodeBench, and BigCodeBench, demonstrating capability on par with GPT-4o. In particular, it proved its excellent performance by scoring 73.7 points on Aider, a code repair benchmark.

To support developers with varying resource environments, the model sizes are provided in fine-grained tiers.

• Support for a total of 6 model sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B
• Both Base models for fine-tuning and Instruct models ready for immediate conversation are released

Key capabilities are as follows:

• Diverse language support: Supports over 40 programming languages, showing particular strength in Haskell and Racket, among others.
• Code reasoning and repair: Excels at reasoning ability to predict code execution processes and repair ability to resolve errors.
• Practical utility: Provides performance immediately applicable to real-world tasks in code assistant and Artifacts environments.