Tech & AI, Explained

qwen

7 min read

🇨🇳 China · Qwen / Tongyi Qianwen (Alibaba) — ⛔ DO NOT USE

Status: 🟩 COMPLETE 🟦 LIVING Verdict: ⛔ AVOID — use Claude / ChatGPT / Gemini / Mistral / Cohere instead.

⛔ This entry exists to explain what Qwen is and why the encyclopedia recommends against using it. The full geopolitical reasoning lives in vendors-chinese-avoid.md.

Front-matter facts

FieldValue
VendorAlibaba Cloud (Hangzhou, Zhejiang, China) — a subsidiary of Alibaba Group
Country / origin🇨🇳 China — mainland PRC
Recommended for Australian users?⛔ NO — not recommended; the consumer chat and Alibaba Cloud API send data to PRC infrastructure; subject to Chinese state-access law and content regulation
Why we recommend againstMainland China-based; subject to PRC Cybersecurity Law, Data Security Law, National Intelligence Law; outputs filtered per CAC regulations; weaker privacy guarantees; Alibaba parent also operates extensive consumer-data businesses in China
Western alternatives by capabilityChat → Claude 🟩 🟦 or ChatGPT 🟩 🟦 or Gemini 🟩 🟦 · Multilingual chat → Mistral Le Chat 🟩 🟦 or Cohere 🟩 🟦 · Open-weight models → Llama 🟥 or Mistral open weights 🟥 or Gemma 🟥 · Image → Imagen 🟥 or Firefly 🟥
First releasedTongyi Qianwen launched April 2023; Qwen-7B (open weights) August 2023; chat product (qwen.ai / tongyi.aliyun.com) ongoing
Last reviewed2026-06-26
Official site(not linking — pick a Western alternative)

What it is (factually)

Qwen (pronounced “chwen,” short for “Tongyi Qianwen” — 通义千问 — roughly “truth-seeking through a thousand questions”) is Alibaba Cloud’s family of large language models, plus the consumer chat product built on them.

It comes from one of China’s largest tech conglomerates:

  • Alibaba Group — Chinese e-commerce giant (Taobao, Tmall, AliExpress, Alibaba.com), payments (Alipay), logistics (Cainiao), and cloud (Alibaba Cloud / Aliyun)
  • Alibaba Cloud — China’s largest cloud provider, also operates in Singapore, Southeast Asia, and (controversially) provides infrastructure to numerous regimes

The Qwen model family includes:

  • Qwen-Max / Qwen-Plus / Qwen-Turbo — closed commercial chat models
  • Qwen-VL — multimodal (vision + language)
  • Qwen2 / Qwen2.5 / Qwen3 / Qwen3.5 — successive generations
  • Qwen Open weights — Alibaba has aggressively released open-weight versions (Qwen-7B, Qwen-14B, Qwen-72B, Qwen2.5-Coder, Qwen2-Audio, etc.)
  • Qwen Image — Alibaba’s image gen
  • Qwen-Video — video gen

Like DeepSeek, Qwen models are technically strong on benchmarks. The open weights are widely downloaded and form a significant chunk of the open-source AI ecosystem (especially in coding and multilingual tasks).

Why we recommend against it

Same five-reason framework as DeepSeek (see vendors-chinese-avoid.md for the full reasoning):

  1. Data goes to China. Consumer chat at qwen.ai / tongyi.aliyun.com and direct Alibaba Cloud API calls route to mainland China servers.
  2. Chinese law gives state agencies access. PRC Cybersecurity Law, Data Security Law, National Intelligence Law all apply. Alibaba cannot refuse data demands from PRC authorities and cannot disclose the requests.
  3. Outputs are politically filtered. Required under CAC generative AI regulations. Independent testing has confirmed Qwen refuses or evades on Tiananmen, Taiwan, Xinjiang, Tibet, Hong Kong, Falun Gong, Xi Jinping criticism, etc.
  4. Privacy guarantees are weaker. Alibaba’s privacy posture is rated weaker than Western providers by independent audits. The parent company operates extensive consumer-data businesses (Alipay financial data, Taobao shopping behaviour, etc.) — the corporate group is fundamentally data-monetisation-oriented.
  5. You don’t gain anything Western providers don’t offer. Qwen is strong, but Claude / ChatGPT / Gemini / Mistral / Cohere all match or exceed it for almost every use case, without these downsides.

Specific concerns particular to Alibaba / Qwen:

  • Alibaba Cloud infrastructure ban from US government use since 2024; Australian Department of Home Affairs has issued cautions; UK NCSC has issued advisories
  • Alibaba operates surveillance products in China (City Brain, smart-city systems) — the corporate culture is not data-minimisation-oriented
  • Alibaba was sanctioned by Chinese regulators in 2021 (the “anti-monopoly” crackdown) — corporate independence from the CCP is meaningfully constrained
  • Qwen’s training-data sources are not disclosed in the way Western frontier labs publish; reproducibility / training-data audit is not possible

”But the open weights are okay, right?”

Same nuance as DeepSeek. Qwen open weights (Qwen2.5, Qwen3, Qwen-Coder, Qwen-Audio etc.) can be downloaded and run on your own hardware OR on Western clouds — which avoids the “data goes to China” problem.

If you genuinely want Qwen’s capabilities via open weights:

  • ✅ Run via Together AI, Fireworks AI, Hugging Face Inference, AWS Bedrock (where available), Groq, or locally via Ollama / LM Studio
  • ✅ Your data does NOT go to China this way
  • ❌ The political-filtering training is still baked into the weights
  • ❌ You can’t audit the training data

Encyclopedia recommendation: prefer Western open-weight alternatives (Llama 🟥, Mistral open weights 🟥, Gemma 🟥, OpenAI gpt-oss 🟥) which cover essentially every use case without the political-filtering and training-trust concerns.

Qwen-Coder is genuinely strong on coding benchmarks, but so are Claude (via Claude Code), GPT-5 (via OpenAI Codex / Cursor), Gemini (via Antigravity / Code Assist), and Codestral (Mistral). No need to compromise.

See open-weights-vs-closed.md 🟥 for the deeper treatment.

What to use instead

If you want to use Qwen for…Use this Western alternative
Multilingual chat (its strongest claim)Mistral Le Chat 🟩 🟦 (excellent on European/Asian languages) · Cohere Aya 🟥 (multilingual research models) · Claude 🟩 🟦 (strong on most major languages)
Coding via open weightsMistral Codestral 🟥 · Llama Code 🟥 · DeepSeek-Coder via Western hosts (if you must — but Mistral / Llama preferred)
Image genImagen 🟥 · Adobe Firefly 🟥 · Midjourney 🟥
Video genSora 🟥 · Veo 🟥 · Runway 🟥
Cheap general APIGemini 3 Flash 🟥 · Claude Haiku 4.5 🟩 🟦 · nano 🟥 · Mistral Small 🟥

Gotchas specific to Qwen

  • “Qwen” branding sometimes appears in unexpected places — Alibaba pushes Qwen models into Aliyun Cloud, DingTalk (Alibaba’s enterprise chat), Quark browser (Alibaba’s mobile browser). If you use Alibaba consumer products, you’re already exposed to Qwen.
  • The English-language UI on qwen.ai or international tongyi.aliyun.com is polished — UI language doesn’t change server location or corporate jurisdiction.
  • Some Western AI tools accidentally bundle Qwen models — when you use a generic “open-weight model” on a self-hosting platform, double-check the model identifier. Qwen is fine to research on Western infrastructure but you should KNOW that’s what you’re using.
  • Alibaba’s M6 / “Tongyi” branding ecosystem includes other AI products beyond chat — Tongyi Wanxiang (image), Tongyi Tingwu (audio), Tongyi Lingma (code) — same parent, same concerns, sometimes branded slightly differently in international markets.
  • Hugging Face hosting of Qwen weights doesn’t help safety concerns — the weights themselves carry the training biases regardless of where you download them from. (Hugging Face is a US company; hosting isn’t the issue. The model’s training is.)
  • DingTalk (Alibaba’s enterprise chat with Qwen built in) is sometimes adopted by Australian businesses with Chinese partners. Be aware that adopting DingTalk inside your business adopts Qwen’s data-handling model for those conversations.
  • Alibaba Cloud international (Singapore-headquartered Alibaba Cloud Intelligence) is sometimes presented as separate from China-based operations. Legally, it’s still ultimately part of the Alibaba Group; PRC law applicability is a topic of regulatory contention rather than a settled “this is safe” answer.

See also

Sources

Graph View

Backlinks