Qwen2.5-Coder Series: Powerful, Diverse, Practical

Qwen2.5-Coder is initialized based on the Qwen2.5 foundation model and is a "powerful," "diverse," and "practical" open-source model. This series of models has been continuously trained with 5.5T tokens of data, including source code, mixed-text code, and synthetic data, resulting in significant improvements in core tasks such as code generation, code reasoning, and code repair.

Powerful: Qwen2.5-Coder-32B-Instruct has become the current state-of-the-art (SOTA) open-source code model, with code capabilities on par with GPT-4o. It demonstrates strong and comprehensive coding abilities while also exhibiting solid general-purpose and mathematical skills.

Diverse: Last month, we open-sourced the 1.5B and 7B versions. This time, we are also releasing the 0.5B, 3B, 14B, and 32B versions. As of now, Qwen2.5-Coder covers six mainstream model sizes, catering to the needs of different developers.

Practical: We explore the practicality of Qwen2.5-Coder in two scenarios, including code assistants and Artifacts, with some examples showcasing the potential applications of Qwen2.5-Coder in real-world scenarios;

This release of Qwen2.5-Coder includes a full series of models in six sizes: 0.5B, 1.5B, 3B, 7B, 14B, and 32B, with both Base and Instruct models available for each size. The Base models are available for fine-tuning by developers, while the Instruct models are ready-to-use official alignment models. All Qwen2.5-Coder models achieve the best model performance (SOTA) at each corresponding size.

The newly released flagship model, Qwen2.5-Coder-32B-Instruct, has set new record scores for open-source models across more than ten mainstream code generation benchmarks, including EvalPlus, LiveCodeBench, and BigCodeBench. It also outperforms GPT-4o on nine benchmarks, such as Aider (which tests code repair ability) and McEval (which evaluates multi-language capabilities), achieving a breakthrough where open-source models surpass closed-source models.

In terms of code reasoning, Qwen2.5-Coder-32B-Instruct has set a new best record for open-source models on the CRUXEval-O benchmark. Additionally, it excels across over 40 programming languages, achieving the highest score among all open and closed-source models on the McEval benchmark, and winning the open-source champion title on the MdEval benchmark, which tests multi-language code repair capabilities.

The full series of Qwen2.5-Coder is open-sourced, making it adaptable to more application scenarios. Whether on the edge or in the cloud, AI large models can better assist developers in programming tasks. Even "beginners" in programming can use the built-in Qwen2.5-Coder code assistant and visualization tools to generate various applications, such as websites, data charts, resumes, and games, through natural language conversations.

As of now, Qwen2.5 has open-sourced over 100 large language models, multimodal models, mathematical models, and code models, with nearly all models achieving the best performance at equivalent sizes. By the end of September, the number of derivative models based on the Qwen series exceeded 74,300, surpassing the 72,800 derivative models of the Llama series. Tongyi Qianwen has thus become the largest generative language model family in the world.

Community

Qwen2.5-Coder Series: Powerful, Diverse, Practical

Related Links

Read previous post:

Alibaba Cloud Community

You may also like

Comments

Alibaba Cloud Community

Related Products

Tongyi Qianwen (Qwen)

Alibaba Cloud for Generative AI

AI Acceleration Solution

Platform For AI