Qwen3-Coder Release Notes

Qwen has announced Qwen3-Coder, a high-performance code model designed for agentic tasks. The flagship variant, Qwen3-Coder-480B-A35B-Instruct, is a 480B-parameter Mixture-of-Experts (MoE) model with 35B active parameters that sets new state-of-the-art results among open models for agentic coding, tool-use, and browser-use, performing comparably to Claude Sonnet 4.

Technical Architecture and Pre-Training

Qwen3-Coder utilizes a Mixture-of-Experts architecture to balance total parameter count with computational efficiency. The pre-training process focused on three primary scaling dimensions:

  • Token Scaling: The model was trained on 7.5 trillion tokens, with a 70% code ratio, to ensure strong coding capabilities without sacrificing general knowledge or mathematical reasoning.
  • Context Window Scaling: Qwen3-Coder natively supports a 256K token context window, which can be extended to 1 million tokens using YaRN. This capacity is optimized for handling repository-scale data and dynamic inputs such as Pull Requests.
  • Synthetic Data Scaling: The team used Qwen2.5-Coder to clean and rewrite noisy training data, improving the overall quality of the pre-training corpus.

Post-Training and Reinforcement Learning

Qwen3-Coder's capabilities are driven by two distinct reinforcement learning (RL) strategies focused on execution and long-horizon tasks.

Execution-Driven Code RL

Rather than focusing solely on competitive programming, Qwen3-Coder was trained using large-scale RL on a broad set of real-world coding tasks. By automatically scaling test cases for diverse coding tasks, the model improved its code execution success rates and general performance across other tasks.

Long-Horizon Agent RL

To solve complex software engineering tasks (such as those in SWE-Bench), the model underwent "Agent RL" training. This process encourages the model to engage in multi-turn interactions—planning, using tools, and incorporating feedback—to solve problems. To support this, Qwen built a scalable infrastructure on Alibaba Cloud capable of running 20,000 independent environments in parallel. Consequently, Qwen3-Coder achieves state-of-the-art performance among open-source models on SWE-Bench Verified without requiring test-time scaling.

Tooling and Integration

Qwen has introduced several ways to integrate Qwen3-Coder into development workflows:

Qwen Code CLI

Qwen Code is a research-purpose command-line interface (CLI) tool forked from Gemini Code and adapted with customized prompts and function calling protocols specifically for Qwen3-Coder.

Third-Party Tool Support

Qwen3-Coder is compatible with several popular developer tools:

  • Claude Code: Users can integrate Qwen3-Coder via a proxy API or the claude-code-router and claude-code-config npm packages.
  • Cline: The model can be configured in Cline by selecting the "OpenAI Compatible" provider and using the Dashscope API endpoint.

API Access and Future Directions

Qwen3-Coder is available via the Alibaba Cloud Model Studio API. The qwen3-coder-plus model identifier is used for API calls through an OpenAI-compatible interface.

Future development for Qwen3-Coder includes the release of additional model sizes to reduce deployment costs and the exploration of self-improvement capabilities for the coding agent to handle increasingly complex software engineering tasks.

Sources