Alibaba has introduced Qwen3-Coder, its most powerful agentic AI model to date, aimed at supporting complex software development tasks. Designed to improve how developers write, manage and debug code, the model leverages a large-scale Mixture-of-Experts (MoE) architecture and is now available as open-source software.
Built for real-world development, Qwen3-Coder can generate new code, handle intricate programming workflows and resolve bugs across large codebases. With the release, Alibaba aims to make software engineering more efficient, accessible and autonomous for developers worldwide.
Enhanced performance for large-scale development
The full model, named Qwen3-Coder-480B-A35B-Instruct, consists of 480 billion parameters but activates only 35 billion per token, allowing for high performance with greater efficiency. Trained on a vast collection of code and general text data, it natively supports a 256,000-token context window, extendable to 1 million tokens—enabling it to process massive codebases in a single run.
Qwen3-Coder also introduces several technical advances in model training, including token scaling, long-context optimisation and synthetic data use. Post-training enhancements such as long-horizon reinforcement learning, or agent RL, allow the model to handle multi-step problem-solving through interaction with external tools. These improvements contribute to Qwen3-Coder’s strong results on SWE-Bench Verified, a benchmark used to measure AI models’ ability to fix real-world software issues. Notably, it achieved state-of-the-art results among open-source models without requiring any additional tuning at test time.
Expanding accessibility through tools and interfaces
To support real-world use, Alibaba is also releasing Qwen Code, a command-line interface (CLI) that allows developers to assign engineering tasks to the model using natural language. Optimised with built-in prompts and interaction protocols, Qwen Code helps users unlock the full capabilities of Qwen3-Coder in day-to-day development.
The model is also compatible with external interfaces, including the Claude Code interface, which offers developers even greater flexibility in integrating Qwen3-Coder into their workflows.
With agentic coding becoming more widely adopted, Qwen3-Coder aims to bridge the gap between autonomous programming tools and practical development environments. Its design reflects the increasing need for AI models that not only understand code but also interact intelligently with software tools and systems.
Global rollout and integration into Alibaba Cloud
Qwen3-Coder-480B-A35B-Instruct is now available on Hugging Face and GitHub. Developers can also access it through Alibaba’s Qwen Chat platform or via cost-effective APIs offered through Model Studio, the company’s generative AI development suite.
The Qwen-based coding models have collectively surpassed 20 million downloads globally. Alibaba Cloud’s AI coding assistant, Tongyi Lingma, will also be upgraded to include Qwen3-Coder’s enhanced features. Since launching in June 2024, Tongyi Lingma’s AI Programmer tool has been used to generate more than 3 billion lines of code, supporting tasks such as code completion, optimisation, debugging, snippet search and automated unit test generation.
With Qwen3-Coder, Alibaba is making a strong push into the AI-for-developers space, offering scalable, open-source tools that reflect ongoing shifts towards more intelligent and autonomous coding solutions.
