The Qwen team has introduced the Qwen2.5-Coder series, a new set of code-specific language models derived from the Qwen2.5 architecture. The series includes six model sizes ranging from 0.5B to 32B parameters and is trained on over 5.5 trillion tokens of cleaned source code and synthetic data.

  • The models achieve state-of-the-art performance across more than 10 benchmarks for code generation, completion, reasoning, and repair.
  • Qwen2.5-Coder-32B-Instruct matches the coding capabilities of GPT-4o while retaining strong general and mathematical skills.
  • The architecture incorporates special tokens to support Fill-in-the-Middle (FIM) operations and repository-level context management.
  • The release includes permissive licensing to facilitate wider adoption by developers.

The authors aim to advance research in code intelligence and support real-world applications through these open-source models.