Unlocking Qwen3‑Coder: The New Open‑Source AI Model Redefining Agentic Coding

Just now, the latest AI programming model Qwen3‑Coder from Tongyi has been officially open‑sourced.

The new Qwen3‑Coder model boasts outstanding code and agent abilities, achieving SOTA results on open‑source benchmarks for Agentic Coding, Agentic Browser‑Use, and foundational coding tasks.

Qwen3‑Coder is released in multiple sizes; the most powerful version, Qwen3‑Coder‑480B‑A35B‑Instruct , is a Mixture‑of‑Experts model with 480 billion parameters, 35 billion active parameters, native 256K context, and can be extended to 1 million tokens via YaRN.

To help developers use Qwen3‑Coder, Tongyi provides the Qwen Code CLI tool, which fully exploits the model’s agentic programming potential. The model’s API can also be combined with tools such as Claude Code and Cline.

Qwen3‑Coder is available on ModelScope, Hugging Face, and will soon be integrated into Alibaba’s AI programming product Tongyi Lingma; its API is already live on Alibaba Cloud Bailei.

Qwen‑Coder Technical Highlights

Data Scaling : 7.5 TB of data (70 % code) to enhance programming ability while preserving general and mathematical skills.

Context Expansion : Native 256K context, extendable to 1 M with YaRN, optimized for repository‑level and dynamic data such as Pull Requests.

Synthetic Data Expansion : Uses Qwen2.5‑Coder to clean and rewrite low‑quality data, significantly improving overall data quality.

Post‑Training

The team applies large‑scale reinforcement learning on real software‑engineering tasks (e.g., SWE‑Bench), enabling long‑horizon RL where the model iteratively plans, selects tools, receives feedback, and makes new decisions. An environment‑scaling system runs 20 k independent environments on Alibaba Cloud, achieving SOTA results on SWE‑Bench Verified.

Qwen Code CLI Installation

Ensure Node.js 20+ is installed, then run:

curl -qL https://www.npmjs.com/install.sh | sh

npm i -g @qwen-code/qwen-code

Or install from source:

git clone https://github.com/QwenLM/qwen-code.git</code><code>cd qwen-code && npm install && npm install -g

Set environment variables for OpenAI‑compatible SDK:

export OPENAI_API_KEY="your_api_key_here"</code><code>export OPENAI_BASE_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"</code><code>export OPENAI_MODEL="qwen3-coder-plus"

After setup, simply type qwen to start coding with Qwen‑Code.

Integration with Claude Code

Two integration methods are provided:

Replace Anthropic base URL with DashScope endpoint:

export ANTHROPIC_BASE_URL=https://dashscope.aliyuncs.com/api/v2/apps/claude-code-proxy</code><code>export ANTHROPIC_AUTH_TOKEN=your-dashscope-apikey

Use claude-code-config for custom routing:

npm install -g @musistudio/claude-code-router</code><code>npm install -g @dashscope-js/claude-code-config</code><code>ccr-dashscope

Then run ccr code to start using Claude Code with Qwen3‑Coder.

Demo Showcase

Various demos created with Qwen3‑Coder include chimney demolition, fireworks animation, typing speed test, rotating sphere, solar‑system simulation, and a duet game.

API Call Example

import os</code><code>from openai import OpenAI</code><code></code><code>client = OpenAI(</code><code>    api_key=os.getenv("DASHSCOPE_API_KEY"),</code><code>    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",</code><code>)</code><code></code><code>prompt = "Help me create a web page for an online bookstore."</code><code></code><code># Send request to qwen3-coder-plus model</code><code>completion = client.chat.completions.create(</code><code>    model="qwen3-coder-plus",</code><code>    messages=[</code><code>        {"role": "system", "content": "You are a helpful assistant."},</code><code>        {"role": "user", "content": prompt}</code><code>    ],</code><code>)</code><code></code><code># Print the response</code><code>print(completion.choices[0].message.content.strip())

Future Outlook

The Tongyi team continues to improve Coding Agent performance, aiming to handle more complex software‑engineering tasks and explore self‑improving agents, while additional model sizes are in development to balance effectiveness and deployment cost.