Alibaba has unveiled Qwen3-Coder, its most sophisticated open-source AI coding model yet, designed to handle complex, multi-step software development tasks with minimal human intervention. The model pushes the boundaries of autonomous coding through a powerful agentic AI architecture that combines scalability, contextual awareness, and real-world programming capabilities.

Designed for Agentic AI Coding at Scale

Built on a Mixture-of-Experts (MoE) architecture, Qwen3-Coder-480B-A35B-Instruct activates only 35 billion of its total 480 billion parameters per token—offering computational efficiency while maintaining high performance. The model delivers state-of-the-art (SOTA) results across benchmarks focused on agentic coding, browser interactions, and tool use.

Trained on a vast corpus of code and general text, Qwen3-Coder natively supports a 256,000-token context window—extendable up to one million tokens—making it capable of processing entire codebases in a single prompt.

Real-World Impact and Tools for Developers

As part of the launch, Alibaba has also open-sourced Qwen Code, a command-line interface tool that allows developers to assign engineering tasks to the AI using natural language. The tool is built to fully harness Qwen3-Coder’s agentic capabilities, thanks to its optimised prompts and interaction protocols.

The model also integrates with the Claude Code interface, offering developers an additional pathway to streamline coding workflows.

“Agentic AI coding is transforming software development by enabling more autonomous, efficient, and accessible programming workflows,” Alibaba said in its release.

Strong Benchmark Performance and Global Adoption

Qwen3-Coder leads among open-source models on SWE-Bench Verified, a key benchmark assessing AI performance on real-world software issues. The achievement comes without relying on inference-time scaling techniques, thanks to advanced post-training approaches like long-horizon reinforcement learning (agent RL).

The broader Qwen model family has already surpassed 20 million downloads worldwide. Its integration with Alibaba Cloud’s coding assistant Tongyi Lingma—which since its June 2024 debut has generated over 3 billion lines of code—is expected to further boost its adoption. Tongyi Lingma will soon be upgraded with Qwen3-Coder’s enhanced features, including improved code completion, debugging, and batch test generation.

Availability

Developers can access Qwen3-Coder-480B-A35B-Instruct on Hugging Face and GitHub, or experiment with it via Qwen Chat and Model Studio, Alibaba’s AI development platform that offers cost-effective API access.