Alibaba releases Qwen3-Coder

Alibaba’s Qwen team released Qwen3-Coder, an open-weight model built for agentic coding tasks — writing, running and iterating on code with tool access rather than producing snippets in a single turn. The flagship variant, Qwen3-Coder-480B-A35B, is a mixture-of-experts model with 480 billion total parameters, of which 35 billion are active for any given token, trained on 7.5 trillion tokens with a code ratio of roughly 70%. It natively supports a 256,000-token context window, extending to 1 million tokens with extrapolation methods, and was released under the permissive Apache 2.0 licence.

Alibaba said the model set new state-of-the-art results among open models on agentic coding, agentic browser use and agentic tool use, and described its performance as comparable to Anthropic’s Claude Sonnet 4 — a claim resting on Alibaba’s own benchmark selection rather than independent evaluation. Alongside the model, the company released Qwen Code, a command-line agent tool, and made the model available through Alibaba Cloud’s Model Studio and compatible with existing agent frameworks such as Claude Code and Cline.

The release extended a pattern set by DeepSeek’s R1 earlier that year: Chinese labs shipping open-weight models, under licences permitting commercial use, that claimed rough parity with the best closed frontier systems on specific tasks. For coding specifically, it added pressure on both open- and closed-weight rivals by targeting the exact workflow — an agent operating a terminal and browser to complete multi-step software tasks — that Western labs were simultaneously building into their own flagship products, including OpenAI’s ChatGPT Agent released five days earlier.