Qwen3-Coder: Agentic Coding in the World
- Source
- qwenlm.github.io
- Author
- Qwen Team
- Date
Context
Qwen's Qwen3-Coder was trained for coding tasks that unfold across multiple interactions with tools. Its training setup let the model plan actions, use tools, receive feedback and decide what to do next. The July 2025 release describes reward-based training in software environments, extending the task from generating code to pursuing a result through successive attempts.
The team also released Qwen Code, a command-line interface that connects the model to coding tools. These are distinct parts of the workflow: the model generates decisions and code, while the interface enables interaction with the working environment. The reported training supports this iterative approach; it does not establish that every repository task can be completed successfully.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
“Qwen3-Coder-480B-A35B-Instruct sets new state-of-the-art results among open models on Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use, comparable to Claude Sonnet 4.”
Checking sign-in…
Loading comments…