The Qwen team has announced the release of Qwen3-Coder-480B and Qwen3-Coder-35B Instruct, marking a significant advancement in open agentic code models and enhancing coding capabilities.
The new models are designed to offer robust architectural design and practical utility for software development tasks. Qwen3-Coder-480B stands as the largest open code model to date, demonstrating state-of-the-art performance across various code benchmarks. Its capabilities encompass code completion, editing, and generation. A notable achievement for Qwen3-Coder-480B is its exceptional performance in code completion, particularly in the Fill-in-the-Middle task, where it achieved a Pass@1 score of 42.0%. The model’s extensive training regimen involved over 3 trillion tokens sourced from diverse datasets, which contributes to its enhanced ability to comprehend and generate code effectively. This vast training data and the model’s sheer size position it as a highly versatile tool for developers, capable of handling a wide spectrum of tasks from fundamental code completion to intricate code generation.
Complementing the 480B model, Qwen3-Coder-35B Instruct is specifically optimized to function as a coding agent, providing assistance to developers with complex coding requirements. This model incorporates advanced features, including a 32K token context window. It has been trained with specialized techniques that facilitate the seamless use of external tools and function calls. This functionality allows developers to invoke external APIs directly within their coding environment, thereby improving productivity and streamlining development workflows. The Qwen3-Coder-35B Instruct model is engineered for robustness and efficiency, making it highly suitable for agent applications that involve complex interactions.
Both Qwen3-Coder-480B and Qwen3-Coder-35B Instruct are now publicly accessible on leading platforms such as Hugging Face and ModelScope. In a move to foster innovation and collaboration within the AI-driven software development community, the Qwen team has open-sourced all associated code, model weights, and comprehensive documentation. This open-source approach is intended to empower researchers and developers to utilize these models, contributing to their further improvement and broader adoption.
