What is Qwen3-Coder
Qwen3-Coder is our most agentic code model to date, representing a breakthrough in AI-powered software development.
🚀 Qwen3-Coder-480B-A35B-Instruct
A powerful 480B-parameter Mixture-of-Experts model with 35B active parameters, supporting 256K context length natively and up to 1M tokens with extrapolation methods.
🎯 State-of-the-Art Performance
Sets new benchmarks among open models on Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use, comparable to Claude Sonnet 4.
🔧 Advanced Training
Trained on 7.5T tokens (70% code ratio) with large-scale reinforcement learning, excelling in coding while preserving general and math abilities.
🌐 Universal Integration
Works seamlessly with community's best developer tools including Qwen Code, Claude Code, and Cline for agentic coding tasks.
⚡ Real-World Focus
Optimized for repo-scale tasks and dynamic data like Pull Requests, enabling true "Agentic Coding in the World".
"As a foundation model, we hope it can be used anywhere across the digital world — Agentic Coding in the World!"
Qwen3-Coder: Advanced AI model architecture for agentic coding tasks
Qwen 3 Coder Core Features
Advanced Pre-Training at Scale
Qwen3-Coder pushes the boundaries of pre-training with massive scale improvements across multiple dimensions:
- 7.5T Tokens: Trained on 7.5 trillion tokens with 70% code ratio, excelling in coding while preserving general and math abilities
- Extended Context: Native 256K context support, expandable to 1M tokens with YaRN for repo-scale tasks
- High-Quality Data: Leveraged Qwen2.5-Coder to clean and rewrite noisy data, significantly improving overall data quality
Large-Scale Code Reinforcement Learning
Revolutionary approach to code generation using execution-driven reinforcement learning:
- Hard to Solve, Easy to Verify: Scaled up Code RL training on diverse real-world coding tasks
- Automatic Test Scaling: Created high-quality training instances by automatically scaling test cases
- Execution Success: Significantly boosted code execution success rates across all tasks
Long-Horizon Agent Reinforcement Learning
Multi-turn interaction capabilities for real-world software engineering tasks:
- 20,000 Parallel Environments: Built scalable system using Alibaba Cloud infrastructure
- Multi-Turn Interaction: Planning, tool usage, feedback processing, and decision making
- SWE-Bench Excellence: State-of-the-art performance on SWE-Bench Verified without test-time scaling
Seamless Tool Integration
Works perfectly with the best developer tools in the community:
- Qwen Code: Research-purpose CLI tool adapted from Gemini CLI with enhanced parser support
- Claude Code: Full compatibility with Claude Code through DashScope API integration
- Cline Support: Easy configuration for OpenAI-compatible API usage
- Universal Foundation: Can be used anywhere across the digital world
Performance Benchmarks
How to Use Qwen3-Coder
Get started with Qwen3-Coder using your favorite development tools. Choose from three popular options:
Option 1: Qwen Code CLI
Research-purpose CLI tool adapted from Gemini CLI with enhanced parser support
Install Node.js 20+
curl -qL https://www.npmjs.com/install.sh | sh
Install Qwen Code
npm i -g @qwen-code/qwen-code
Or install from source:
git clone https://github.com/QwenLM/qwen-code.git
cd qwen-code && npm install && npm install -g
Configure Environment
export OPENAI_API_KEY="your_api_key_here"
export OPENAI_BASE_URL="https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
export OPENAI_MODEL="qwen3-coder-plus"
Start Coding
qwen
Now enjoy your vibe coding with Qwen-Code and Qwen!
Option 2: Claude Code Integration
Use Qwen3-Coder with Claude Code for seamless development experience
Install Claude Code
npm install -g @anthropic-ai/claude-code
Configure Proxy API (Option A)
export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/api/v2/apps/claude-code-proxy
export ANTHROPIC_AUTH_TOKEN=your-dashscope-apikey
Or Configure Router (Option B)
npm install -g @musistudio/claude-code-router
npm install -g @dashscope-js/claude-code-config
ccr-dashscope
Then start using: ccr code
Option 3: Cline Configuration
Configure Qwen3-Coder with Cline for enhanced development workflow
Configuration Steps:
- Go to the Cline configuration settings
- For API Provider, select 'OpenAI Compatible'
- For the OpenAI Compatible API Key, enter the key obtained from Dashscope
- Check 'Use custom base URL' and enter:
https://dashscope-intl.aliyuncs.com/compatible-mode/v1
- Enter model name: qwen3-coder-plus
🎯 Quick Start Recommendation
New to Qwen3-Coder? We recommend starting with Qwen Code CLI for the best out-of-the-box experience. It's specifically optimized for Qwen3-Coder with enhanced parsing and tool support.
Frequently Asked Questions about Qwen3-Coder
What makes Qwen3-Coder different from other code models?
Qwen3-Coder is our most agentic code model to date, featuring a 480B-parameter Mixture-of-Experts architecture with 35B active parameters. It supports 256K context length natively (expandable to 1M tokens) and sets new state-of-the-art results on Agentic Coding, Browser-Use, and Tool-Use benchmarks, comparable to Claude Sonnet 4.
How was Qwen3-Coder trained?
Qwen3-Coder was trained on 7.5T tokens with 70% code ratio, using advanced techniques including large-scale reinforcement learning, synthetic data cleaning with Qwen2.5-Coder, and long-horizon RL (Agent RL) across 20,000 parallel environments. This training approach follows the "Hard to Solve, Easy to Verify" principle for optimal code generation.
What tools can I use with Qwen3-Coder?
Qwen3-Coder works seamlessly with multiple developer tools: Qwen Code (our research CLI tool adapted from Gemini CLI), Claude Code (via DashScope proxy API), and Cline (with OpenAI-compatible configuration). All tools connect through DashScope's API endpoints for consistent performance.
How do I get API access to Qwen3-Coder?
You can access Qwen3-Coder through Alibaba Cloud's DashScope platform. Simply request an API key from the Model Studio platform, then configure your environment with the base URL: https://dashscope-intl.aliyuncs.com/compatible-mode/v1 and model name: qwen3-coder-plus.
What programming languages does Qwen3-Coder support?
Qwen3-Coder excels across multiple programming languages and frameworks. With its 70% code training ratio from diverse real-world coding tasks, it supports popular languages like Python, JavaScript, TypeScript, Java, C++, Go, Rust, and more. It's particularly strong at repo-scale tasks and understanding complex codebases.
What is the context length capability of Qwen3-Coder?
Qwen3-Coder natively supports 256K tokens context length, which can be extended up to 1M tokens using YaRN extrapolation methods. This makes it ideal for repo-scale tasks, analyzing large codebases, and handling dynamic data like Pull Requests for true "Agentic Coding in the World".
How does Qwen3-Coder perform on coding benchmarks?
Qwen3-Coder achieves state-of-the-art performance among open-source models on SWE-Bench Verified without test-time scaling. It sets new benchmarks on Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use tasks, with performance comparable to Claude Sonnet 4 while being an open foundation model.
Can I use Qwen3-Coder for commercial projects?
Yes! Qwen3-Coder is designed as a foundation model that can be used anywhere across the digital world. Please check the specific license terms and usage policies on the DashScope platform for detailed commercial usage guidelines and any applicable restrictions.
What makes Qwen3-Coder "agentic"?
Qwen3-Coder is trained with long-horizon RL (Agent RL) to engage in multi-turn interactions involving planning, tool usage, feedback processing, and decision making. This enables it to solve real-world software engineering tasks through autonomous agent-like behavior, rather than just generating code snippets.
How do I get started with Qwen Code CLI?
Getting started is simple: 1) Install Node.js 20+, 2) Run npm i -g @qwen-code/qwen-code, 3) Configure your DashScope API key and environment variables, 4) Type qwen to start coding! The CLI tool is specifically optimized for Qwen3-Coder with enhanced parsing and function calling protocols.