DeepSeek, Qwen, GLM, and Kimi are Chinese-developed AI model families. You may see their names in AI apps or as model options inside other tools.
There are two main ways an animal advocate might “use a Chinese model”:
- Use the company’s own app. For example, you could open DeepSeek’s consumer app and chat with it directly.
- Choose the model inside a tool you already use. You might select it in Cursor, have an agent such as OpenClaw use it, or choose it for an n8n automation.
For everyday advocacy work, we see no good reason to use a Chinese company’s own chat app. Use a paid frontier assistant instead. This means one of the leading options, such as Claude or ChatGPT. Our guide to picking the right AI tools for your team explains why.
Chinese-developed models are more relevant in the second situation. They can be cheap and capable, so one may suit a specific automation or technical task. They can also bring extra risks in some workflows.
The model and the service are different
An AI model is a piece of software. The company that creates it does not have to be the company running it.
DeepSeek, for example, publishes downloadable model files on Hugging Face under the MIT License. A U.S. company can download one of these models and run it on its own servers.
If Cursor or an n8n automation connects to that version, your prompt goes to the separate company hosting the model. It does not necessarily go to DeepSeek the company. Check which company is hosting the model before using it for organizational work.
The host’s country matters too. If a company in China hosts the model, assume Chinese authorities could require it to provide stored chats or other data. DeepSeek’s privacy policy, for example, says it collects prompts and chat history, stores personal data in China, and may share data to comply with government requests. This does not mean authorities routinely read every chat. It does mean you should keep sensitive organizational information out of a Chinese-hosted service.
When might a Chinese-developed model make sense?
| Possible benefit | What to check |
|---|---|
| Lower cost for a high-volume task | Compare the cost of each successful result after retries and corrections. Some models use many more tokens while working through an answer. You are charged for those extra tokens, so a model with a cheaper token price can still cost more per task. |
| Better performance on a specific task or language | Test it on examples from your real work. Benchmark scores may not reflect your task. |
| A choice of companies to host the model | Open models may be available through several hosting companies. An organization with enough technical infrastructure can also run one on its own servers, although this is unlikely to be practical for a small nonprofit. |
| Less dependence on one U.S. model developer | Another model family gives you an alternative if your usual model changes or becomes unavailable. Test it before you need it, and keep a stronger model available for difficult cases. |
Cheap prices are not unique to Chinese-developed models. As of September 2026, OpenAI prices GPT-5.6 Luna at $0.20 per million input tokens and $1.20 per million output tokens. OpenAI designed it for cost-sensitive, high-volume work. Test it alongside any Chinese-developed model you are considering. Our guide to finding smart, low-cost AI models gives a short testing process.
Test security and sensitive topics
Models can behave differently even when benchmark scores look similar. A NIST evaluation found security and censorship problems in the DeepSeek versions it tested. They were more vulnerable to agent hijacking and gave inaccurate answers on some political topics.
Those findings do not apply automatically to every Chinese-developed model or every later version. They do justify extra testing when a workflow handles untrusted web content or covers China, politics, investigations, or other sensitive subjects.
Use only providers approved for the kind of data in the workflow. Keep donor records, employee information, unreleased investigations, whistleblower material, legal advice, and credentials out of an unapproved AI tool.
Am I legally allowed to use Chinese models?
Usually, yes. As of September 3, 2026, we found no blanket rule stopping an ordinary nonprofit from using a Chinese-developed model. This is general information, not legal advice, and rules differ by country and contract.
The narrow U.S. restriction most relevant to this question covers the DeepSeek app or service on intelligence-community national security systems. It also covers systems run by contractors or other organizations on behalf of the intelligence community. The law does not ban ordinary nonprofit use.
DeepSeek’s own terms of use say that mainland Chinese law governs the service and that legal disputes may go through courts where DeepSeek is registered. Using its app creates that direct legal relationship with the company. This is another reason we recommend Claude or ChatGPT for everyday advocacy work.
Your organization may still face restrictions through a government contract, grant agreement, insurer, client agreement, or its own AI policy. Ask a lawyer when a contract, regulated data, or government work makes the answer unclear.
Our recommendation
Use Claude or ChatGPT as your normal AI app. Consider a Chinese-developed model inside an existing tool or automation only when testing shows a clear cost or performance benefit. Confirm who runs it and use it only with suitable data. Keep a tested fallback.
