Alibaba unveils Qwen3-Coder open-source AI model for coding tasks

Alibaba has introduced Qwen3-Coder, its latest open-source agentic artificial intelligence (AI) coding model for software development.

The Qwen3-Coder-480B-A35B-Instruct model, which features a Mixture-of-Experts (MoE) architecture, is engineered for agentic AI tasks such as code generation, managing complex workflows, and debugging entire codebases. The model is open-sourced and available for download on platforms including Hugging Face and GitHub.

Model architecture

The Qwen3-Coder-480B-A35B-Instruct model contains a total of 480 billion parameters but only activates 35 billion parameters per token, resulting in more efficient computing without compromising performance. The model's architecture is designed to deliver high efficiency for development tasks, scaling resources according to need.

According to Alibaba, Qwen3-Coder achieves competitive results across various industry-standard benchmarks. "Qwen3-Coder-480B-A35B-Instruct achieves competitive results against leading state-of-the-art (SOTA) models across key benchmarks," the company stated.

Open source tools and integration

In addition to the model itself, Alibaba is releasing Qwen Code, a command-line interface (CLI) tool developed to help users delegate programming tasks to AI using natural language. This tool is optimised with custom prompts and interaction protocols to support the agentic programming capabilities of Qwen3-Coder in real applications.

The statement highlights, "Alibaba is open-sourcing Qwen Code, a powerful command-line interface (CLI) tool that enables developers to delegate engineering tasks to AI using natural language. Optimized with custom prompts and interaction protocols, Qwen Code unlocks the full potential of Qwen3-Coder for real-world agentic programming. The model also supports integration with the Claude Code interface, making it even easier for developers to execute their coding tasks."

Qwen3-Coder is built on a large set of code and general text data, providing a foundation for robust agentic coding tasks. It natively supports a context window of 256K tokens, extensible up to 1 million tokens, which allows the model to process and analyse substantial codebases within a single session.

Training and benchmark performance

Performance improvements stem from scaling across tokens, context length, and synthetic data during pre-training, as well as post-training techniques including long-horizon reinforcement learning (agent RL). This enables the model to solve complex problems through multi-step interactions with external tools.

This advancement allows the model to solve complex, real-world problems through multi-step interactions with external tools. As a result, Qwen3-Coder achieves SOTA performance among open-source models on SWE-Bench Verified (a benchmark for evaluating AI models' ability to solve real-world software issues), even without test-time or inference scaling.

Applications and compatibility

Qwen3-Coder is designed to enable agentic AI coding, which Alibaba describes as a mechanism for more autonomous and accessible programming workflows. With its open-source status, agentic coding features, and compatibility with widely-used developer tools and interfaces, the model is intended for global adoption by the software development community.

The company noted, "Agentic AI coding is transforming software development by enabling more autonomous, efficient, and accessible programming workflows. With its open-source availability, strong agentic coding capabilities, and seamless compatibility with popular developer tools and interfaces, Qwen3-Coder is positioned as a valuable tool for global developers in software development."

Developers can access the Qwen3-Coder-480B-A35B-Instruct model via Qwen Chat or through cost-aware APIs on Model Studio, Alibaba's generative AI development platform. To date, Qwen-based coding models have been downloaded more than 20 million times worldwide.

Product enhancements

Alibaba also announced that Tongyi Lingma, its Qwen-powered coding assistant, is scheduled for an upgrade with Qwen3-Coder's new agentic capabilities. Since its introduction, the assistant's "AI Programmer" feature has produced over 3 billion lines of code and provides several functions such as code completion, optimisation, debugging, snippet search, and automated unit test generation.

"Tongyi Lingma, Alibaba Cloud's Qwen-powered coding assistant, will soon be upgraded with Qwen3-Coder's enhanced agentic capabilities. Since its launch in June 2024, Tongyi Lingma's 'AI Programmer' feature - offering code completion, optimization, debugging support, snippet search, and batch unit test generation - has generated over 3 billion lines of code," the company explained.