Alibaba Launches 480B Parameter AI Coding Model Qwen3-Coder

HANGZHOU, China — Alibaba Group announced the release of Qwen3-Coder, an artificial intelligence model designed for software development tasks.

The model, called Qwen3-Coder-480B-A35B-Instruct, contains 480 billion parameters and activates 35 billion parameters per token, according to the company. The system is built on a Mixture-of-Experts architecture.

The company stated that the model can generate code, manage coding workflows, and debug across codebases. 

The system supports a context window of 256,000 tokens, extendable to one million tokens, the company said.

The model was trained using long-horizon reinforcement learning techniques and a dataset of code and text data, Alibaba said.

The company stated that the model outperformed existing models on SWE-Bench Verified, a software issue resolution benchmark.

Alibaba also released Qwen Code, a command-line interface tool for assigning engineering tasks to the AI system using natural language.

The model supports integration with the Claude Code interface, the company said.

The model is available on the Hugging Face and GitHub platforms. Access is also provided through Qwen Chat and application programming interfaces via Model Studio, Alibaba’s AI development platform.

According to Alibaba, Qwen-based coding models have been downloaded more than 20 million times. 

The company’s coding assistant, Tongyi Lingma, has generated over three billion lines of code since its launch in June 2024.

Tongyi Lingma provides code completion, optimization, debugging support, snippet search, and unit test generation, according to the company. 

Alibaba plans to upgrade Tongyi Lingma with capabilities from the new model.

The model is open-sourced, according to the company.