This Qwen Code review 2026 covers the open-source terminal agent from Alibaba’s Qwen team — what it actually does, what it costs after the free tier disappeared, and where it falls short. This review is based on official documentation, pricing pages, and community reports — we did not run the tool hands-on.

If you arrived here searching for a free Claude Code alternative, that framing is out of date. Alibaba discontinued Qwen Code’s OAuth free tier on 2026-04-15, and the option no longer appears in the /auth dialog (Qwen Code docs). What is left is a fast-moving Apache-2.0 CLI with an unusually long feature list, a paid entry point, and a few sharp edges.

Qwen Code review 2026: the short verdict

Qwen Code is a capable, unusually feature-rich terminal agent with a broad model allowlist and a genuine Apache-2.0 licence. It scores 7.9/10 here. The catch is commercial: the free tier is gone, the paid entry point is $50 a month, and the project still carries a heavy open-issue backlog.

What Qwen Code actually is

Qwen Code is the open-source terminal coding agent built by the Qwen team at Alibaba. Its own repository describes it as “the open-source AI coding agent for your terminal, editor, desktop, browser, and chat.” It is not the Qwen3-Coder model, which is a separate product reviewed elsewhere on this site.

The public repository showed 28,060 stars, 3,077 forks and 1,498 open issues on 2026-09-22, under an Apache-2.0 licence, with TypeScript as the primary language and a creation date of 2025-06-26 (github.com/QwenLM/qwen-code). The current npm release is 0.24.3, requiring Node.js 22 or newer (npm registry). If the model behind it is what interests you, our Qwen3-Coder 480B review covers that separately.

Installing Qwen Code: one command, then restart your terminal

Official docs list a Linux/macOS curl script, a global npm install, Homebrew, and a Windows PowerShell one-liner. Node.js 22 or newer is required for npm. Restart your terminal after installing so PATH changes take effect, then start a session by running qwen inside your project folder.

curl -fsSL https://qwen-code-assets.oss-cn-hangzhou.aliyuncs.com/installation/install-qwen-standalone.sh | bash
npm install -g @qwen-code/qwen-code@latest
brew install qwen-code

Those three commands come straight from the official docs. Beyond the terminal, the project ships a VS Code extension, Zed and JetBrains integrations, a desktop app for macOS, Windows and Linux, an experimental Web UI started with qwen serve --open, GitHub Actions support, daemon mode over HTTP and SSE, a TypeScript SDK, and alpha Python and Java SDKs.

Pricing after the free tier shut down

Qwen Code no longer offers a free tier. The OAuth option was discontinued on 2026-04-15 and cannot be selected in the /auth dialog any more. Today you pick between an Alibaba ModelStudio plan, a built-in third-party provider key, or a custom OpenAI-, Anthropic- or Gemini-compatible endpoint such as a self-hosted proxy, Ollama or vLLM.

Third-party trackers reported that the OAuth daily quota had already been cut from 1,000 requests to 100 on 2026-04-13, two days before shutdown; that figure is tracker-reported rather than vendor-stated, and the official docs simply describe the provider as discontinued (Qwen Code docs). The current /auth menu offers three top-level paths: Alibaba ModelStudio (Coding Plan, Token Plan or Standard API Key), Third-party Providers with a built-in list covering DeepSeek, Grok, MiniMax, Z.AI, Kimi, Idealab, ModelScope, OpenRouter and Requesty, and Custom Provider for any compatible endpoint.

The paid headline is the Alibaba Cloud ModelStudio Coding Plan Pro at $50/month, with three caps that apply at the same time: up to 6,000 requests per five hours, up to 45,000 requests per week, and up to 90,000 requests per month. Whichever limit is reached first pauses the service (Alibaba Cloud ModelStudio pricing).

Quota mechanics matter too. The five-hour bucket rolls forward minute by minute, the weekly cap resets Monday at 00:00 UTC+08:00, and the monthly cap resets on your renewal date. Slots restock daily at 00:00 UTC+08:00 and availability is not guaranteed — the plan is capacity-limited (Alibaba Cloud ModelStudio pricing). The cheaper Lite plan closed to new subscriptions on 2026-03-20, and renewals and upgrades stopped on 2026-04-13 (Alibaba Cloud ModelStudio pricing).

Why 6,000 requests per five hours is not 6,000 chats

This is the most misunderstood number in the product. One query consumes quota per model call, not per conversation: a simple task typically burns five to ten calls, and a complex task ten to thirty or more. So 6,000 requests is a call budget a coding agent can work through faster than the label suggests.

The Coding Plan also fixes your model choice. The documented Pro allowlist is qwen3.7-plus, qwen3.6-plus, kimi-k2.5 (all three vision-capable), glm-5, MiniMax-M2.5, qwen3.5-plus, qwen3-max-2026-01-23, qwen3-coder-next, qwen3-coder-plus and glm-4.7; models outside that list are not supported. Endpoints are regional — Beijing uses https://coding.dashscope.aliyuncs.com/v1 and international uses https://coding-intl.dashscope.aliyuncs.com/v1 — and the wrong base URL means your usage is not deducted from the plan (Alibaba Cloud ModelStudio pricing).

How Qwen Code scores

We rate Qwen Code 7.9/10 overall, built from five sub-scores: Ease of Use 7, Features 9, Performance 8, Documentation 8 and Support 6. Features lead because the documented list is unusually broad; Support trails because the open-issue backlog is large and the free tier is gone.

How we tested

This is desk research. We did not install Qwen Code, run a session, or benchmark it on any machine. Every claim here comes from the official documentation, the public GitHub repository, the vendor pricing page, or clearly-labelled community and third-party tracker reports, and each is linked inline.

Concretely, that means no installation, no timed runs, no side-by-side coding trials, and no measured token costs. Where a figure comes from a third-party tracker rather than Alibaba, we say so. Where a feature is documented but unverified by us, we describe it as documented rather than confirmed. Our AI tools directory lists everything covered on this site under the same standard.

Features: what the documentation lists

The documented feature set covers Auto-Memory, Auto-Skills, SubAgents, Agent Teams, MCP, Hooks, Skills, Rules, Memory, Approval Mode, LSP integration, token caching, resident context, cost tracking, sandboxing, plugins, worktrees, code review, Agent Arena, scheduled tasks, computer use, browser use, structured output, internationalisation, status line and dual output.

SubAgents deserve a note because they behave differently from a normal prompt. They run as independent assistants with their own system prompt, separate conversation history, controllable tool access and autonomous execution. The inner claude-code and codex subagents delegate to separately installed native tools, need a trusted workspace, are unavailable in safe mode, and do not fall back to Qwen’s own model if the executable is missing. Chat channels — Telegram, WeChat, DingTalk, WeCom, Feishu and QQ Bot — start with qwen channel start, and the team publishes weekly product updates going back to 02-03.

Configuring providers in ~/.qwen/settings.json

Configuration lives in ~/.qwen/settings.json. A modelProviders entry maps a provider id to an array of model configs, each requiring an id and optionally an envKey. Credentials are read from the environment and never persisted in the settings file. The built-in provider auth types are openai, anthropic, gemini and vertex-ai.

The silent-skip trap in migrated settings files

In an already-migrated $version: 4 settings file, a modelProviders entry still using the older wrapped {protocol, models} shape is skipped silently — no error, no warning, the model simply never appears. Keep the array bare. Headless and CI runs can use BAILIAN_CODING_PLAN_API_KEY, OPENAI_BASE_URL and OPENAI_MODEL, or the qwen -p "..." flag.

Qwen Code vs Claude Code, Codex CLI and Gemini CLI

This is positioning, not benchmarking. Qwen Code is Apache-2.0 and model-agnostic across an allowlist that includes non-Qwen models. Claude Code and Codex CLI are tied to their own vendors’ models and subscriptions, and Gemini CLI follows the same pattern. Qwen Code’s difference is surface breadth and a bring-your-own-endpoint escape hatch.

Kilo Code is the closest open-source comparison point, and our Aider vs OpenHands vs Claude Code breakdown covers the wider field. The practical distinction is that Qwen Code lets you point it at a self-hosted server, a proxy, Ollama or vLLM through the Custom Provider path, so you are never locked to one vendor’s inference bill — while Claude Code, Codex CLI and Gemini CLI each ship their own subscription model that you should check on their own pages. For a side-by-side table, see compare AI coding tools.

Who should not use Qwen Code

Skip it if you need a free tier, because none exists. Skip it if you want a slow-moving, heavily supported product — 1,498 open issues on a repository created in mid-2025 points the other way. Skip it if you need mature Python or Java SDKs, or a stable web interface.

More specifically, avoid Qwen Code if any of these apply to you:

• You will not pay $50/month and have no third-party key or self-hosted endpoint to bring (Alibaba Cloud ModelStudio pricing).
• You need the closed Lite plan, which stopped accepting subscribers on 2026-03-20 (Alibaba Cloud ModelStudio pricing).
• You need guaranteed plan availability; slots restock daily at 00:00 UTC+08:00 and are not guaranteed (Alibaba Cloud ModelStudio pricing).
• You rely on daemon mode, the Web UI, or the Python and Java SDKs at production quality — they are marked experimental or alpha.
• You want a free choice of model on a Coding Plan; the allowlist is fixed.

Bottom line

Qwen Code is one of the broadest open-source terminal agents available, and the Apache-2.0 licence plus custom-provider support make it genuinely flexible. It is also young, paid at the entry point, and carrying a large issue backlog. Try the bring-your-own-key route before committing to a plan.

For a beginner, the honest sequence is: install it, connect a cheap third-party key, and see whether the feature depth justifies the migration once your usage grows. If it does not, our DeepSeek Harness DSH review and our AI code review tools roundup cover cheaper starting points.

📊 See how Qwen Code compares →

📖 Related Reads

• CodeIntel Log — code quality, debugging, and software engineering benchmarks
• ToolBrain — tool reviews, LLM comparisons, and AI workflow guides

Cross-links automatically generated from None.