Qwen Code CLI: Getting Started Guide for AI Coding 2026

What Is Qwen Code CLI?

Qwen Code is an open-source command-line AI agent maintained by QwenLM, Alibaba's model team. It lives in your terminal, reads your codebase, edits files, runs shell commands, and ships features end to end. The QwenLM/qwen-code GitHub repo is Apache-2.0 licensed and built around the Qwen3-Coder family of models.

Architecturally it started as a fork of Gemini CLI and kept the same interactive TUI shape, but the prompts and function-calling protocols are tuned for Qwen3-Coder. If you have used Claude Code or Gemini CLI before, the muscle memory transfers. /help, @file.ts references, slash commands, and a session-based memory model all work the way you expect.

The headline numbers worth knowing: 256K native context, ~1M with YaRN extrapolation, and a steady climb on GitHub through Q1 2026 according to the public repo star history. For a non-Western model CLI that is a strong adoption signal, and it is the main reason I added it to my rotation alongside the ant CLI and Antigravity CLI.

How to Install Qwen Code

The fastest path is npm. You need Node.js 20 or higher; older versions silently fail with module resolution errors that look unrelated. I learned that one the hard way on a Debian server still on Node 18.

macOS and Linux (npm)

# Confirm Node.js >= 20
node --version

# Install Qwen Code globally
npm install -g @qwen-code/qwen-code

# Verify
qwen --version
# qwen-code 0.4.x

macOS and Linux (official install script)

If you do not want to manage Node yourself, the official curl script bundles Node and the CLI together:

bash -c "$(curl -fsSL https://qwen-code-assets.oss-cn-hangzhou.aliyuncs.com/installation/install-qwen.sh)"

Windows

On Windows, run the official .bat installer from an Administrator PowerShell or cmd window:

curl -fsSL -o %TEMP%\install-qwen.bat https://qwen-code-assets.oss-cn-hangzhou.aliyuncs.com/installation/install-qwen.bat && %TEMP%\install-qwen.bat

WSL2 with Ubuntu 22.04 also works and is what I use on my Windows box. The npm path inside WSL is usually less painful than the native Windows installer.

Authenticating After the OAuth Shutdown

Here is the part most existing 2026 guides still get wrong. The original Qwen OAuth free tier - the one where you ran qwen and a browser opened to log in with your Alibaba account - was discontinued on 2026-04-15. If you try to follow a guide from February or March, you will hit a 401 on first prompt and assume the install is broken. It is not.

You now need an API key. Three paths work today, in order of how often I reach for them.

Path 1 - Alibaba Cloud Model Studio (DashScope)

The official default. Generate a key at the Model Studio Qwen Code coding plan page, then drop it into ~/.qwen/settings.json:

{
  "apiKey": "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxx",
  "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
  "model": "qwen3-coder-plus"
}

Path 2 - OpenAI-Compatible Provider (OpenRouter, Together, Fireworks)

If your team cannot egress to Alibaba Cloud, point Qwen Code at any provider that hosts a Qwen3-Coder variant on an OpenAI-compatible endpoint. OpenRouter is the easiest because it brokers across multiple backends and you pay one bill:

{
  "apiKey": "sk-or-v1-xxxxxxxxxxxxxxxxxxxxxxxx",
  "baseUrl": "https://openrouter.ai/api/v1",
  "model": "qwen/qwen3-coder"
}

Path 3 - Self-Hosted vLLM or Ollama

Qwen3-Coder weights are open. If you have a GPU with enough VRAM (the 7B variant runs comfortably on a single 24GB card; the flagship Mixture-of-Experts build needs serious hardware), serve it with vLLM and point Qwen Code at localhost:

{
  "apiKey": "not-needed",
  "baseUrl": "http://localhost:8000/v1",
  "model": "Qwen/Qwen3-Coder-7B-Instruct"
}

The CLI does not care that the key is fake; it just needs the header present. This is the cheapest long-term setup if you already run inference hardware.

First Run - Launching qwen in Your Project

cd into a repo and type qwen. The interactive TUI opens, shows the working directory, and waits for a prompt. There is no project setup step. The CLI picks up your settings from ~/.qwen/settings.json and starts a session.

cd ~/code/my-next-app
qwen

# Inside the TUI:
> Summarize the architecture of @src/app and flag any dead code.

A few keystrokes I use every session:

• /help - list all slash commands
• /clear - reset the conversation but keep settings loaded
• @path/to/file - inline a file into your prompt
• /quit - exit cleanly (Ctrl+C also works)

On my first real task - asking it to refactor a 380-line Next.js route handler into smaller server actions - Qwen3-Coder Plus produced a working diff in about 22 seconds end to end. Latency is the number I care about most for interactive use, and that is competitive with Claude Sonnet 4.6.

Configuring Models and the 1M Context Window

The model field in settings.json is the main lever. Qwen3-Coder ships in a few sizes, and your provider decides which IDs are available. Common ones in mid-2026:

• qwen3-coder-plus - DashScope flagship hosted variant, what I default to
• qwen3-coder-flash - faster, cheaper, smaller, good for routine edits
• qwen/qwen3-coder - OpenRouter alias for the open weights
• Qwen/Qwen3-Coder-7B-Instruct - the size people self-host most often

About the 1M context claim. Qwen3-Coder is trained at 256K. YaRN positional extrapolation can stretch it to roughly 1M tokens, and that is the number Alibaba quotes. I have measured this in practice and recall starts dropping somewhere past 400K tokens, which matches what other practitioners are seeing. Keep the YaRN flag off for normal coding work; only flip it on for tasks like full-repo summarization where you need to dump a few hundred files into a single prompt.

Qwen Code in CI/CD (Non-Interactive Mode)

The -p flag is what makes Qwen Code useful in pipelines. Pass a prompt, the CLI runs it once, prints the result, and exits. No TUI, no readline, no surprises. Here is the minimal GitHub Actions workflow I use for automated PR descriptions:

name: Qwen PR Summary

on:
  pull_request:
    types: [opened, synchronize]

jobs:
  summarize:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      pull-requests: write
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - uses: actions/setup-node@v4
        with:
          node-version: '20'

      - run: npm install -g @qwen-code/qwen-code

      - name: Generate PR summary
        env:
          QWEN_API_KEY: ${{ secrets.QWEN_API_KEY }}
        run: |
          mkdir -p ~/.qwen
          cat > ~/.qwen/settings.json <<EOF
          {
            "apiKey": "$QWEN_API_KEY",
            "baseUrl": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
            "model": "qwen3-coder-flash"
          }
          EOF
          git diff origin/${{ github.base_ref }}...HEAD \
            | qwen -p "Summarize this diff for a PR description. Use bullet points." \
            > summary.md

      - name: Post summary as comment
        uses: marocchino/sticky-pull-request-comment@v2
        with:
          path: summary.md

If you are running this on a public repo, harden the workflow first. I wrote up the pattern - tool allowlists, egress blocks, script caps - in Harden Claude Code GitHub Actions. The same defenses apply to Qwen Code because the failure mode (an injected comment hijacks the agent) is identical.

Qwen Code vs Claude Code vs Gemini CLI

I run all three. They are not interchangeable. Here is the honest breakdown after about six weeks of using Qwen Code as my secondary CLI.

Verdict. Claude Code still wins on multi-step reasoning, plan-then-execute workflows, and complex refactors. Qwen Code wins when (a) the open-weights license matters, (b) you need cost control on high-volume runs, or (c) you have on-prem GPUs sitting idle. Gemini CLI is being sunset on 2026-06-18 - I covered the migration path to agy in detail here.

Common Errors and Fixes

Frequently Asked Questions