Back to Gap Map Product / UX / Orchestration & agents

Qwen Code

Alibaba Cloud
open source / Overall score: 4.4(strong)

Qwen Code is a coding agent from Alibaba's Qwen team that runs in the terminal and through editor plugins, a desktop app, a browser UI and chat apps. It works with the OpenAI, Anthropic, Gemini and Qwen APIs or a local model, and ships sub-agents, agent teams, MCP, plan mode and SDKs for three languages. The project began as a fork of Google's Gemini CLI and now develops on its own.

The root LICENSE names Google LLC beside Qwen as copyright holder, carried over from Gemini CLI 0.8.2, the code the project was forked from. Qwen Code stopped syncing with upstream at its 0.1 release, so the two are read as separate products.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public
core features withheld
no — the tree at 0df419d carries one root Apache-2.0 LICENSE, Apache-2.0 or MIT files in a few packages, and no ee/, enterprise/, commercial/ or proprietary/ path. The paid options the project's terms list are model access through a Qwen account or an Alibaba Cloud Coding Plan key

Qwen Code is released under Apache-2.0, and its public repository holds the whole agent along with its SDKs, so it builds and runs from published source. The project calls itself an open-source tool, and signing in with a Qwen account or an Alibaba Cloud Coding Plan key pays for model access rather than for any part of the agent.

Adoption

3 medium confidence
3.0

Adoption is the trailing month of npm downloads of @qwen-code/qwen-code, the package that is the CLI itself, and the Homebrew package installs the same npm tarball. The README leads with a standalone installer that fetches a prebuilt archive from Alibaba Cloud storage or GitHub releases and falls back to the npmmirror registry. None of those installs reach the npm count, so the figure is a lower bound on use.

Capability

5 medium confidence
5.0

Qwen Code covers the same ground as OpenHands and as Gemini CLI, the agent it was forked from, with sub-agents and agent teams, MCP, plan mode, a sandbox, computer use and many model providers, plus editor, desktop, web and chat front ends and SDKs in three languages. Its SWE-bench Verified results come from the Qwen Code team running its own model rather than from an independent board, the same limit that applies to the scores OpenHands publishes.

  • https://raw.githubusercontent.com/QwenLM/qwen-code/main/docs/evaluation.md recorded 2026-10-19

    The evaluation page compares seven Qwen Code versions on SWE-bench Verified "using the same Qwen 3.7 Max model and configuration", with average scores from 76.40% to 77.80% over 500 cases and three trials per version.

  • https://raw.githubusercontent.com/QwenLM/qwen-code/main/README.md recorded 2026-10-19

    The README's feature table checks "SubAgents, Agent Teams, Dynamic Workflows", "MCP, Plan Mode, LSP Integration", "Auto Mode, Sandbox, Git Worktrees" and "Computer Use (desktop automation)", and lists "Multi-protocol — Supports OpenAI, Anthropic, Gemini, and Qwen APIs", a headless mode, a qwen serve daemon and three SDKs.

Verified 2026-10-19, except some scores, still awaiting re-confirmation