Alibaba has unveiled its top-level agentic AI coding model 'Qwen3-Coder'. This model is designed for high-performance software development and demonstrates excellent performance in code generation, complex coding workflow management, and debugging the entire codebase.
Qwen3-Coder-480B-A35B-Instruct is based on a Mixture-of-Experts (MoE) architecture, activating 35 billion parameters out of a total of 480 billion parameters while providing efficiency without performance degradation. This model has recorded achievements comparable to state-of-the-art (SOTA) models in key benchmarks such as agentic coding, browser usage, and tool usage.
Additionally, Alibaba offers an open-source powerful interface (CLI) tool called Qwen Code, enabling developers to delegate engineering tasks to AI using natural language. This tool is optimized with customized prompts and interaction protocols to maximize the capabilities of Qwen3-Coder, and it integrates with the Claude Code interface, allowing developers to perform coding tasks more easily.
Qwen3-Coder supports a context window of 256,000 tokens and can be extended up to 1 million tokens. This model expands the number of tokens, context length, and scale of synthetic data during the training process, and it applies innovative techniques in subsequent training processes to enhance actual software problem-solving abilities.
This model changes the paradigm of software development, enabling efficient and autonomous programming workflows. Qwen3-Coder is offered as open-source and boasts high compatibility with major development tools and interfaces. It is currently available on Hugging Face and GitHub, allowing developers to leverage the model through cost-efficient expense APIs.
Qwen3-Coder has already recorded over 20 million downloads worldwide, and Alibaba Cloud's Qwen-based coding assistant 'Tongyi Lingma' is expected to be upgraded with enhanced features of Qwen3-Coder.