Is it safe to run DeepSeek or Qwen locally?
Short answer: Three things get mixed up: the hosted app, the model file, and the software that runs it. Here is what each one does and what to check.
Apply for a place From $10,000 setup, then $48,000 a year. Five places a week.
Can one answer cover DeepSeek and Qwen running locally?
No single answer covers it, because people use one name for three different things: the hosted app, the model file, and the program that runs the file. This page does not give a safe or unsafe verdict. It tells you what each piece does and what you should check.
The three pieces have different owners, different terms and different places where your data can go. Mixing them up is how people end up worried about the wrong thing, or relaxed about the wrong thing.
What is the hosted app, and what does its policy say?
The hosted app is the website or phone app run by the company. You type a prompt, it travels to the company's servers, and the company's policy governs what happens next.
As of November 2026, DeepSeek's privacy policy, updated February 10, 2026, says its hosted service collects, processes and stores personal data in the People's Republic of China. It also says inputs are used to train its models, with an opt-out by emailing the company. Read the current policy yourself, because these terms change.
Press reports say some US government bodies have restricted the hosted app on official systems. That concerns the hosted app on official systems. It does not describe running the open files on your own machine, and it does not describe private businesses.
What is the open weights file?
The open weights file is the model itself, a large file of numbers ("weights") that you can download and run on your own computer. Downloading it is a different act from using the company's app, so the hosted policy above is a separate matter.
A program has to load the file before you can use it, which brings us to the third piece.
What does the software that runs the file do?
The runner is the program that loads the file and answers your prompts. Ollama is a common one. According to Ollama's own FAQ, when you run a model locally the company does not see your prompts or data, and its cloud models are a separate feature that you can switch off.
Whether anything leaves your machine depends on the runner and the settings you chose, not on the model alone. You can check this yourself: watch outbound network activity while the model works, and turn off any cloud features you do not use. We have not tested any specific setup for this page, so treat it as a checklist, not a result.
For a longer look at the trade-off, read local AI with Ollama versus the cloud.
What licenses do these models use?
Each model carries its own license, and you should read it on the maker's model page before you rely on it. As of November 2026, the makers' pages say the following. Check them again, because licenses change between releases.
Terms differ even inside one family, so never assume that one license covers every size or version. Our guide to local AI model licenses explains what to look for.
- Recent DeepSeek releases (R1 and later V3 versions) are published under the MIT license.
- The original V3 split its code (MIT) from its weights, which use a separate model license that says it supports commercial use.
- Smaller distilled R1 models carry the license of the base model they were built on, such as Apache 2.0 for Qwen-based ones and Llama licenses for Llama-based ones.
- The Qwen3 open models, from the smallest to the largest, use Apache 2.0.
- Older Qwen2.5 used Apache 2.0 for most sizes but not all. The 72B size uses Qwen's own license.
What should you check before you start?
Check four things before you run any open model: where you got the file, which license it carries, which runner you use, and what that runner can reach on your network. Each is quick to look at.
Download from the maker's own page where you can. Keep the license text with the file. Put the machine behind logins, and keep it off the open internet unless you need it there. If you want the model to have no network path at all, that is called an air-gapped setup.
If you want help, what are the options?
You can do this yourself or have a person do it. If you would rather have a person do it, Freeborn Labs sets up and runs a private assistant in a cloud account you own, then trains you to use it. We do not hold your files, and we do not train on customer data.
If you would rather build it yourself, start with the DIY guides. Picking which model to run is its own question, covered in how to choose an open model for private AI.
What are the limits?
Running a model on your own hardware lowers the number of parties that hold your prompts. It does not remove the need for strong passwords, access limits, backups and your own review of every answer, because any model can be wrong or make things up.
This page is general information, not legal, medical, financial or tax advice. Whether a given model or setup fits your duties, contracts or rules is a question for your own counsel, compliance officer or adviser. We promise no outcomes. Find more questions answered in the private AI answers hub.
Questions
Does a government restriction on the DeepSeek app apply to running the files myself?
Press reports describe restrictions on the hosted app on official systems. They concern that app, not open weights you run yourself, and not private businesses. If you work under rules that cover your tools, ask your own counsel or compliance officer.Does downloading an open model send my data to its maker?
The model file is only one piece. What leaves your machine depends on the program that runs it and its settings, so check outbound network activity and switch off cloud features you do not need.Can I use a Qwen model for a business?
Read the license on the exact model's page first. The Qwen3 open models use Apache 2.0, while older Qwen2.5 has one size with Qwen's own license. Terms can change, so check again before each new download.What is a distilled model?
A distilled model is a smaller model trained to imitate a larger one. For the smaller R1 models, the license follows the base model they were built on, so read that license, not only the DeepSeek name.Is the hosted app the same as the open model?
No. The hosted app is a service run by the company under its own policy. The open model is a file you run under a license, with software you choose.Related guides
- How do you choose an open AI model (Llama, Qwen, Mistral and others) to run privately?
- How Do You Check a Local AI Model's License Before Using It?
- Ollama local vs cloud models: what actually leaves your machine?
- What Is an Air-Gapped Local AI Setup?
- Done-for-you private AI setup service
- Run AI without a public chatbot seeing your data
- Private uncensored AI: the full guide
- All guides to getting your own private AI agent
Apply for a place
Takes thirty seconds. No payment now.
Adults only (18+). Private uncensored AI · Terms · Privacy · All guides