Qwen Code

Short answer. Qwen Code is the Qwen team’s open-source AI coding agent (Apache-2.0) for terminal, editor, desktop, browser and chat, built as a multi-protocol, multi-platform TypeScript framework. It is Alibaba’s flagship open coding agent and the most model-agnostic surface in its portfolio — supporting OpenAI, Anthropic, Gemini and Qwen APIs plus DeepSeek, MiniMax, Z.AI, Kimi, OpenRouter and local models.

Key facts.

• Ships as a CLI (npm), Desktop app, VS Code “Qwen Code Companion” (Beta), Web UI and IM channels (Telegram/DingTalk/WeChat/Feishu); originally based on Google Gemini CLI v0.8.2, independent since v0.1.
• Five permission modes, Seatbelt/Docker sandboxing, auto-memory (with QWEN.md as the guaranteed instruction file), subagents, MCP, computer use and Agent Arena for head-to-head model comparison.
• The CLI is free (Apache-2.0); the user pays the model provider directly, and BYO keys work for any supported provider.
• Alibaba billing: international Coding Plan Pro $50/month; China Token Plan (Beijing only) Personal ¥39–¥499/month, team seats ¥150–¥1,398; or pay-as-you-go Model Studio API keys.
• Multi-protocol framework: OpenAI, Anthropic, Gemini and Qwen APIs, plus DeepSeek, MiniMax, Z.AI, Kimi, OpenRouter and local models (Ollama/vLLM).

What this means. Qwen Code’s value is optionality, not a single model — it is the open client that explicitly supports competitors’ models, making it the opposite of Qwen-Agent’s Qwen-native design and complementary to Qoder’s closed multi-model routing.

What is uncertain. The Web UI and daemon (qwen serve) are experimental, the Qwen OAuth free tier was discontinued 2026-04-15, and Auto Mode’s classifier is biased toward blocking (failing closed on outage). Independent benchmark evidence of Qwen Code’s coding quality is not recorded in this database.

Sources.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-agents-qwen-code-1	Qwen Code GitHub repository	https://github.com/QwenLM/qwen-code	Official documentation	—	2026-09-20	high	—
src-agents-qwen-code-2	Qwen Code docs - overview	https://qwenlm.github.io/qwen-code-docs/en/users/overview/	Official documentation	—	2026-09-20	high	—
src-agents-qwen-code-3	Alibaba Cloud Model Studio Coding Plan	https://www.alibabacloud.com/help/en/model-studio/coding-plan	Official documentation	—	2026-09-20	high	—
src-agents-qwen-code-4	Alibaba Cloud Token Plan overview	https://help.aliyun.com/en/model-studio/token-plan-overview	Official documentation	—	2026-09-20	high	—

Why it matters

Qwen Code is the most model-agnostic agent in Alibaba’s portfolio, and the database’s clearest example of a vendor shipping open tooling that works against its own model lock-in. Its multi-protocol TypeScript framework explicitly supports competitors’ models — DeepSeek, Kimi, Z.AI, MiniMax, OpenAI, Anthropic, Gemini — alongside Qwen, with the CLI free and the user paying whatever provider they route to.

China AI Hub analysis indicates Qwen Code matters as Alibaba’s bet that open tooling wins developer mindshare even when those developers ultimately route to other vendors’ models — a bet that mirrors how the Qwen model family itself is open-weight. The agent’s value is the framework (sandboxing, five permission modes, MCP, computer use, Agent Arena), not any single model, which is why it can afford to be genuinely model-optional: Alibaba monetizes the access (Coding Plan, Token Plan, Model Studio keys) rather than the client.

How it differs from Kimi Code and GLM Coding Plan

The three coding agents split on openness and model posture.

• Kimi Code is Moonshot’s first-party open client bound to its own models (Kimi K3/K2.7-Code), monetized through Kimi membership — the tightest integration between tool and model, around K3’s 1M-token output.
• GLM Coding Plan is not a client — it is a subscription that injects GLM models into third-party tools (Claude Code, Codex, Cursor, OpenClaw), monetizing model access inside everyone else’s client.
• Qwen Code is an open multi-protocol client that routes to any provider — model-optional by design, free as a tool, with Alibaba monetizing access through its Coding/Token plans.

China AI Hub analysis: the three are the same question — “how does a Chinese vendor capture coding-agent demand?” — answered three ways. Moonshot binds a free open client to its own models; Zhipu skips the client and sells model access inside other tools; Alibaba gives away the most flexible client and charges for access. Qwen Code is the most open on both axes (open source and model-agnostic), which is precisely why it competes on framework quality rather than model exclusivity.

Practical implications

For developers. Qwen Code is the “bring your own model” open client — free CLI, five permission modes, Seatbelt/Docker sandboxing, MCP, computer use, subagents and Agent Arena for head-to-head model comparison. It runs on npm, desktop, VS Code, Web UI and chat channels, and is pipe-friendly for CI (qwen -p).

For self-hosters. The framework is Apache-2.0 and the sandboxing is real — Seatbelt/Docker — but the default Docker image is intentionally minimal (Java not included), and sandboxing reduces rather than eliminates risk; GUI apps may not run in sandboxes.

For cost-conscious users. The CLI is free and BYO keys work for any provider, so the marginal cost is whatever model you point it at. Alibaba’s own plans (Coding Plan $50/month international; Token Plan ¥39–¥499/month China, Beijing only) are optional, not required.

What the evidence shows

The evidence is strong on the framework and honest about the experimental edges. The GitHub repo and docs document the multi-surface distribution (CLI, desktop, VS Code, Web UI, IM channels), the multi-protocol model support, the five permission modes, Seatbelt/Docker sandboxing, auto-memory/QWEN.md, MCP, computer use and Agent Arena; the Model Studio and Token Plan pages fix Alibaba’s own billing options.

The caveats are explicit rather than hidden: the Web UI and daemon are experimental, the Qwen OAuth free tier ended 2026-04-15, Auto Mode’s classifier is biased toward blocking and fails closed on outage, and sandboxing reduces but does not eliminate risk. China AI Hub analysis indicates this is a mature open project that discloses its own rough edges — the “experimental” and “fails closed” annotations are themselves evidence of production realism, not immaturity, and buyers should read the Auto Mode bias as a deliberate safety posture rather than a defect.

Where this fits

Workload	Relevance
Multi-protocol coding (any provider, local or API)	High
Sandboxed execution (Seatbelt/Docker, 5 permission modes)	High
CI automation and pipe-friendly workflows (qwen -p)	High
Remote agent via chat channels (Telegram/DingTalk/WeChat/Feishu)	High
Multi-model comparison (Agent Arena)	High
Production Web UI / daemon (qwen serve)	Low (experimental)
Independent coding benchmark	No evidence recorded

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

Field reference

Field	Value	Evidence type
Underlying model	Not pinned (multi-protocol: OpenAI, Anthropic, Gemini, Qwen, DeepSeek, MiniMax, Z.AI, Kimi, OpenRouter, local)	Official
Target users	Developers (terminal, editor, desktop, browser, chat)	Official
Platform	CLI (npm), Desktop app, VS Code companion, Web UI, IM channels	Official
OS	Cross-platform (CLI/Desktop); Docker sandbox	Official
Browser / computer use	Browser and Computer Use	Vendor-reported
Coding	Yes (coding agent)	Official
Autonomous task execution	Yes (planning, multi-agent subagents)	Vendor-reported
MCP	Yes	Official
Tool calling	Yes	Official
Memory	Yes (auto-memory, QWEN.md)	Official
Workflow	5 permission modes; Seatbelt/Docker sandboxing; multi-protocol model routing	Official
API	Yes (user pays model provider)	Official
Pricing	CLI free; Alibaba Cloud Coding Plan Pro $50/month or Token Plan ¥39–¥499/month	Vendor-reported
Region	Not publicly documented	Not publicly documented
Open-source	Yes — Apache-2.0	Official
Deployment	Self-hosted	Official
Limitations	Web UI/daemon experimental; sandboxing reduces but not eliminates risk	Official
Source	Qwen Code GitHub	Official
Last verified	2026-09-20	Official

See the Alibaba Cloud company profile, the Qwen-Agent framework, the Qoder platform, the choosing-a-coding-model guide, the Model Studio platform, and the site’s AI agents and MCP technology pages.

Labels used above: Official fact (from the Qwen Code GitHub repo and docs), Vendor-reported claim (pricing and capability statements by Alibaba Cloud), and China AI Hub analysis (our synthesis, always introduced as such). No independent benchmark of Qwen Code’s coding quality is currently recorded.