aistaging/qwen3-coder repository overview

⁠Qwen3-Coder-30B-A3B-Instruct

Qwen3-Coder-30B-A3B-Instruct is a state-of-the-art coding model developed by Alibaba Cloud's Qwen team. This streamlined mixture-of-experts (MoE) model delivers impressive performance in agentic coding, browser automation, and foundational coding tasks while maintaining efficiency through sparse activation. With only 3.3B parameters activated out of 30.5B total, it achieves exceptional performance while remaining computationally efficient.

The model excels at tool calling and function execution, making it ideal for agentic workflows where code generation needs to interact with external tools and APIs. It supports an extensive context window of 256K tokens natively (extendable to 1M tokens using Yarn), enabling repository-scale code understanding and generation. Qwen3-Coder is specifically designed to work seamlessly with platforms like CLINE and features a specially designed function call format for agentic coding scenarios.

This non-thinking mode model generates direct code responses without intermediate reasoning blocks, making it optimized for production environments where clean, immediate code output is preferred. It supports conversational interfaces and can handle complex multi-turn coding dialogues while maintaining context across long interactions.

⁠Characteristics

⁠Using this model with Docker Model Runner

docker model run qwen3-coder

For more information, check out the Docker Model Runner docs⁠.

⁠Architecture

Qwen3-Coder uses a Mixture of Experts (MoE) architecture with 128 expert networks, activating only 8 experts per token. This sparse activation pattern enables the model to maintain a large total parameter count while keeping computational costs manageable through selective activation.

⁠Key Features

Agentic Coding: Built-in tool calling capabilities with a specialized function call format that works across multiple platforms including Qwen Code and CLINE. The model can seamlessly integrate with external tools and APIs.

Long Context Understanding: Native support for 262K token context windows, with extensibility up to 1 million tokens using Yarn technique. This enables comprehensive repository-level code analysis and generation.

Browser Automation: Strong performance on browser-use tasks, enabling automated web interaction and testing workflows.

Conversational Interface: Supports multi-turn conversations with maintained context, making it ideal for interactive coding assistants and pair programming scenarios.

Non-Thinking Mode: Generates direct code output without intermediate reasoning steps, optimized for production use cases requiring clean, immediate responses.

⁠Recommended Settings

For optimal performance, the following parameters are recommended:

• Temperature: 0.7
• Top-p: 0.8
• Top-k: 20
• Repetition Penalty: 1.05
• Max Output Length: 65,536 tokens

⁠Links

• https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct⁠
• https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF⁠
• Official Blog⁠
• GitHub Repository⁠
• Documentation⁠
• Technical Report⁠

⁠Considerations

• Requires transformers>=4.51.0 for proper model loading (earlier versions will result in KeyError: 'qwen3_moe')
• For systems with limited memory, consider reducing the context length from the default 262K to 32K tokens to avoid out-of-memory issues
• The model is designed for non-thinking mode only and will not generate intermediate reasoning blocks
• GGUF quantized versions are available through the unsloth repository for reduced memory footprint and faster inference
• Tool calling functionality requires proper formatting of function definitions and parameters

⁠Generated by

This model card was automatically generated using cagent-action⁠. Want to learn more about Docker Model Runner? Check out the project repository: https://github.com/docker/model-runner⁠.