Alibaba Qwen 3 Coder Open-Supply AI Model Introduced With Agentic Capabilities

Alibaba’s Qwen workforce launched a brand new synthetic intelligence (AI) coding mannequin on Tuesday. Dubbed Qwen 3 Coder, the mannequin comes with a number of agentic capabilities together with agentic coding, agentic browser-use, and agentic tool-use. The researchers have launched just one variant of the mannequin up to now, the Qwen3-Coder-480B-A35B-Instruct, which is essentially the most highly effective variant within the household. In phrases of efficiency in coding, Alibaba’s AI workforce claims the open-source mannequin is similar to Anthropic’s Claude Sonnet 4 mannequin. The new massive language mannequin (LLM) is out there to obtain regionally.

Qwen 3 Coder Model With Agentic Capabilities Released

In a weblog put up, the researchers detailed the brand new agentic coding device. Available in open-source, people can obtain the weights from Qwen’s Hugging Face itemizing and GitHub itemizing. The mannequin is out there with a permissive Apache 2.0 licence which permits for each tutorial and business utilization. Alongside the mannequin, an open-source command-line device dubbed Qwen Code can also be obtainable for agentic coding.

Coming to the mannequin, the Qwen 3 Coder is a mixture-of-expert (MoE) mannequin with 480 billion parameters. Out of that, the mannequin features with 35 billion energetic parameters and a context size of two,56,000 tokens natively. The context window may be expanded to at least one million tokens utilizing extrapolation strategies. The researchers highlighted that the mannequin helps agentic coding, agentic browser-use, and agentic tool-use.

The firm claims that Qwen 3 Coder achieved state-of-the-art (SOTA) efficiency amongst open-source fashions on the SWE-Bench Verified benchmark. Here, SOTA refers back to the highest rating which was beforehand unattainable by some other mannequin. The Qwen workforce stated this rating was achieved by making a scalable system utilizing Alibaba Cloud’s infrastructure that would run 20,000 impartial environments in parallel.

To allow agentic coding, the workforce has additionally launched the Qwen Code command-line device. Built from Gemini Code, it has been geared up with customized prompts and performance calling protocols. These functionalities permit the AI mannequin to not solely write and edit code but additionally to deploy and execute it in an built-in growth atmosphere (IDE).

While Qwen Code natively helps Qwen 3 Coder AI mannequin, it may also be built-in with the OpenAI software program growth package (SDK) when calling LLMs. On the opposite hand, the Qwen coding mannequin may also be used with Claude Code. But builders might want to request an API key on Alibaba Cloud Model Studio platform.