Running Qwen on your own phone is safe from a data standpoint. A downloaded model is a file of numbers. It has no way to send your prompts to Alibaba, to China or anywhere else; whether anything leaves your phone depends on the app running it. The questions worth asking are different ones: how the hosted Qwen service treats your data if you use that instead, and how the model’s training shapes its answers on some political topics.
What is Qwen? #
Qwen is a family of AI models from Alibaba Cloud. In February 2026 Alibaba released Qwen 3.5 as open weights under the Apache 2.0 license, which allows commercial use and modification. According to the Qwen 3.5 model card, the models accept text, images and video, cover 201 languages, and reason step by step by default before they answer.
The small sizes are the ones that fit on phones. Personal LLM offers three:
| Model | Download size | RAM needed | Good for |
|---|---|---|---|
| Qwen 3.5 0.8B | 0.81 GB | 2 GB or more | Very low-end phones; fast, with limited quality |
| Qwen 3.5 4B | 2.74 GB | 3 GB or more | Everyday phones; the recommended first download |
| Qwen 3.5 9B | 5.68 GB | 6 GB or more | The best answers in the catalog, for math, code and careful reasoning |
Can a downloaded AI model send your data to China? #
No. A model in GGUF format, which is what phone apps that run llama.cpp use, holds the model’s weights and some metadata. It’s data, not a program. The software that loads it (here, llama.cpp) does the computing, and the app decides whether anything touches the network.
So judge the app, not the model’s country of origin:
- What does the app connect to? Personal LLM downloads model files once, straight from Hugging Face, and loads ads from Google AdMob in the free version. It has no server of its own, so chats, photos and prompts have nowhere to go.
- Does it work in airplane mode? A true on-device app keeps answering with no connection. Try it.
- Where did the file come from? Get models from the publisher’s official page on Hugging Face, or from an app that does. GGUF files don’t run code the way some older model formats can, but a file from an unknown uploader can still be a badly made or altered model.
Hosted Qwen vs running it yourself #
Alibaba also runs Qwen as an online chatbot, the way OpenAI runs ChatGPT. That’s a different privacy situation.
| Alibaba’s hosted Qwen chat | Qwen running on your phone | |
|---|---|---|
| Where your prompts go | Alibaba’s servers | Nowhere; processed on the phone |
| Who stores your chats | Alibaba, under its privacy policy | Only your phone |
| Rules that apply to the service | The operator’s policies and jurisdiction | Only the laws that apply to you |
| Extra server-side filtering | Up to the operator | None |
| Needs internet | Yes | Only for the one-time download |
| Model size | Large server models | 0.8B to 9B, smaller and less capable |
If you’re uneasy about a Chinese company holding your conversations, the hosted chatbot is where that concern applies. Running the open weights locally removes it.
Is Qwen biased? #
Every language model reflects its training data and the choices its makers made during tuning. American and European models have leanings too, and all of them can be confidently wrong.
Qwen’s particular issue is regulatory. China’s Interim Measures for the Management of Generative AI Services, in force since August 15, 2023, require generative AI services to uphold the Core Socialist Values and bar content judged to undermine the socialist system or national security. Labs based in China build their models under those rules. The practical result is that Chinese-developed models can be cautious, evasive or in line with official positions on some questions about Chinese politics and history.
For most uses you won’t notice. Code, math, recipes, translation, summaries and writing help don’t touch those topics. If your questions do, don’t take any single model’s word for it, Qwen or otherwise.
How to test a model’s bias yourself #
You don’t need a research lab. A phone with two or three models installed is enough:
- Pick 5 to 10 questions on subjects you care about, worded neutrally. Include a couple where you already know the facts well.
- Ask Qwen first. Use the Precise preset (temperature 0.3) so the answers are more consistent.
- Switch models mid-chat. In the app you can change the model from the chat header and ask the same question of Gemma 4 (Google DeepMind) and Ministral 3 (Mistral AI) in the same conversation.
- Regenerate each answer once or twice. If a model takes different positions each time, it isn’t confident either way.
- Turn on thinking mode. Qwen 3.5 shows its step-by-step reasoning in a collapsible panel, which sometimes reveals where it’s steering.
- Ask for the strongest case on each side. A balanced model can argue both; a model that refuses one side is telling you something.
- Check against primary sources. Compare with encyclopedias, official records or reporting from several outlets.
A simple table keeps track of what you find:
| Question | Qwen 3.5 | Gemma 4 | Ministral 3 | Matches sources? |
|---|---|---|---|---|
| Your question | Answered, hedged or refused |
A difference doesn’t prove one model is right. It shows you where to look harder.
What about GLM 4.6V Flash? #
GLM 4.6V Flash comes from Z.ai (formerly Zhipu AI), another China-based lab. It’s a 9-billion-parameter vision model released under the MIT license, good at reading documents, charts and screenshots, and it needs a phone with 8 GB of RAM or more. The same reasoning applies: run locally, it sends nothing anywhere, and on some political topics its answers may reflect the rules its developer works under.
Who should pick a different model? #
- People whose questions often involve Chinese politics or history and who’d rather use a model from a lab outside China. Gemma 4 and Ministral 3 are in the same app.
- Anyone on a work phone where the employer or agency restricts Chinese-developed AI software. Some organizations do, so check your policy.
- Anyone whose older phone runs Qwen 3.5 4B slowly. Gemma 4 E2B (2.04 GB) is built for phones and covers more than 140 languages.
For a broader look at why models refuse different things, see do offline AI chatbots have content filters. If you want to try a Qwen variant that isn’t in the catalog, you can run a custom GGUF model on your phone by pasting its Hugging Face link. And for what stays on the device in general, read how personal data stays protected with on-device AI.
Frequently asked questions #
Does Qwen send data to China? #
The model itself can’t send anything; it’s a data file. Whether data leaves your phone depends on the app. An on-device app with no server sends your prompts nowhere, especially offline. Alibaba’s hosted Qwen chatbot is different: there, your prompts go to Alibaba’s servers.
Is Qwen open source? #
Qwen 3.5’s weights are released under Apache 2.0, so you can download, use and modify them, including commercially. Alibaba hasn’t published its full training data, so “open weights” is the more precise description.
Is Qwen better than Gemma on a phone? #
It depends on the task and the phone. Qwen 3.5 4B is the recommended first download in Personal LLM for its balance of speed and quality, and the 9B gives the best answers in the catalog. Gemma 4 E2B and E4B are built for phones and cover more than 140 languages. The app shows tokens per second on every reply and has a built-in benchmark, so you can compare them on your own hardware.
Can I use Qwen for work? #
Qwen 3.5’s Apache 2.0 license allows commercial use. Some government bodies and employers ban Chinese-developed AI software on their devices, so check your organization’s policy before installing it on a work phone.
Does thinking mode make Qwen less biased? #
No, but it makes the model’s reasoning visible. You can see which facts it weighed and where it chose to hedge, and that makes a slant easier to spot.