Qwen3-Coder: Agentic Coding for the World

Qwen3-Coder-480B-A35B-Instruct is a 480B-parameter Mixture-of-Experts model that uses 35B active parameters. It natively supports 256K tokens, and can be extended up to 1M tokens with extrapolation methods.

Among open models, it achieves new state-of-the-art performance in Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use, with performance described as comparable to Claude Sonnet 4. Alongside it, the CLI tool for agentic coding, Qwen Code, was also open-sourced.

In pretraining, three axes were pushed forward.

• It was trained on 7.5T tokens, of which 70% was composed of code ratio.
• It natively supports 256K context, and was designed to be extendable up to 1M via YaRN.
• Qwen2.5-Coder was leveraged to clean and rewrite noisy data, improving synthetic data quality.

In the post-training stage, code tasks and agent tasks were each reinforced. Code RL applied large-scale reinforcement learning to real coding tasks that can be verified through execution, and automatically scaled test cases to create diverse, high-quality training instances. As a result, code execution success rates improved significantly, and performance gains carried over to other tasks as well.

long-horizon RL (Agent RL), closer to real software engineering, was also introduced. For tasks like SWE-Bench, the model was trained to perform planning, tool use, feedback incorporation, and decision-making across multiple turns, and to support this, a system running 20,000 independent environments in parallel was built on Alibaba Cloud infrastructure.

The tooling was organized together as well. Qwen Code is a research CLI forked from Gemini CLI, with custom prompts and a function calling protocol added to draw out Qwen3-Coder's agentic capabilities. It also provides an OpenAI SDK-compatible API, Claude Code integration, and Cline configuration examples, allowing it to be attached to a variety of development workflows.

The core message is clear. Qwen3-Coder is being positioned not just as a code generator, but as a foundation model targeting agentic development tasks that require execution and tool use, and going forward, more model sizes and directions for self-improvement will continue to be explored.