Alibaba releases Qwen3-Coder and the Qwen Code agent

Background

Alibaba's Qwen team presented Qwen3-Coder as its most agentic code model to date.^[1]

What happened

On 22 July 2025 the Qwen team announced Qwen3-Coder-480B-A35B-Instruct. It is a mixture-of-experts model with 480 billion total parameters and 35 billion active.^[1]

The team released the model alongside Qwen Code, an open-source command-line agent adapted from Gemini CLI.^[1]

Key details

The model has a 256K-token native context, which can be extended to 1 million tokens with YaRN extrapolation.^[1]

Its long-horizon agent reinforcement learning ran 20,000 independent environments in parallel on Alibaba Cloud.^[1]

The team claimed state-of-the-art results among open models on agentic coding, browser use and tool use. It said the model was comparable to Claude Sonnet 4 and reported results on SWE-bench Verified without test-time scaling.^[1]

Several terms in the announcement need explaining. An agent is a model that takes actions, such as calling tools, reading the results and deciding the next step, rather than only producing text. Long-horizon tasks are those that need many steps before they are complete. Reinforcement learning is a training method in which a model improves by attempting tasks and receiving feedback. Running 20,000 environments in parallel means that many such attempts happen at the same time, and each environment is a simulated workspace in which the model can act.^[1]

SWE-bench Verified is a benchmark built from real issues in open-source software projects. A score on it shows how often a model can resolve those issues. Test-time scaling means spending more computation while the model answers. The team said its SWE-bench Verified result was reported without that method.^[1]

The native context is the amount of text a model can read in one go. The 256K-token window can be extended to 1 million tokens with YaRN, a technique that lets a model handle inputs longer than the length it was trained on.^[1]

Impact

The release paired the model with Qwen Code, so developers could use it from a terminal. The model was also said to work with third-party tools such as Claude Code and Cline.^[1]