On July 23, Alibaba Cloud officially announced that its latest AI programming large model, Qwen3-Coder, is fully open-sourced, which quickly sparked a buzz in the intelligent programming field. Qwen3-Coder has achieved top-tier performance among open-source models in Agentic Coding, Agentic Browser-Use, and basic coding tasks, marking a new stage in intelligent programming technology.

The Qwen3-Coder model series offers multiple size options. The first version to be open-sourced is its most powerful one - Qwen3-Coder-480B-A35B-Instruct. This model adopts an advanced MoE architecture, with up to 480B parameters and 35B activated parameters. It natively supports a 256K context length, and can be extended to 1M using YaRN technology, providing strong support for processing large-scale code repositories and dynamic data.

WeChat screenshot_20250723081042.png

In the pre-training phase, the Tongyi team significantly enhanced Qwen3-Coder's coding capabilities through multi-dimensional expansion strategies. In terms of data, the total training data of 7.5T includes as much as 70% code, ensuring that the model maintains general and mathematical abilities while possessing excellent programming capabilities. In terms of context expansion, the model's native long-context processing capability is optimized for repository-level code and dynamic data, greatly improving the efficiency and accuracy of Agentic Coding. Additionally, by using synthetic data expansion technology, low-quality data was cleaned and rewritten using Qwen2.5-Coder, further enhancing the overall data quality.

In the post-training phase, the Tongyi team innovatively adopted a large-scale reinforcement learning strategy driven by execution. By automatically expanding test examples, a large number of high-quality training instances were created. This strategy not only significantly improved the success rate of code execution but also had a positive impact on other tasks. Especially in real-world software engineering tasks such as SWE-Bench, Qwen3-Coder demonstrated excellent self-planning, tool calling, and decision-making capabilities, achieving the best results among open-source models on SWE-bench Verified.

To facilitate developers' use, the Tongyi team also open-sourced the command-line tool Qwen Code. This tool provides enhanced parsers and tool support for the Qwen3-Coder series models, enabling developers to better leverage the model's potential in agent-based programming. At the same time, the API of Qwen3-Coder can be used in collaboration with excellent programming tools such as Claude Code and Cline, offering developers a more flexible and efficient programming experience.

Currently, Qwen3-Coder is fully open-sourced on platforms such as ModelScope Community and HuggingFace, and global developers can download and use it for free. In addition, the model will soon be integrated into Alibaba's AI programming product, Tongyi Lingma, further expanding its application scenarios. The Alibaba Cloud BaiLian platform has also launched the API of Qwen3-Coder, providing developers with a more convenient access method.

ModelScope Community: https://modelscope.cn/models/Qwen/Qwen3-Coder-480B-A35B-Instruct

Hugging Face: https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507

Qwen Code GitHub: https://github.com/QwenLM/qwen-code