Skip to content

DeepSeek, Qwen and Kimi for Business: Where Your Data Goes

DeepSeek, Qwen and Kimi are capable and cheap, and they come from Chinese companies. Here is what each says about where data goes, and how Mistral and self-hosting compare.

7 min read

DeepSeek, Qwen and Kimi for Business: Where Your Data Goes

Someone on your team has tried DeepSeek, or Qwen, or Kimi. It was fast, it was good, and it cost little or nothing. Now they want to use it for client work.

These models are real options. They are also built and hosted by companies in China, and that changes the question you need to answer first. Not "is it any good?" but "where does what we type go, and what happens to it there?"

This guide covers the three, with Mistral and self-hosting as the comparison. It rests on each vendor's own pages and the OAIC's privacy guidance. It is not legal advice. Your obligations depend on your industry, your contracts and the kind of data involved, so check with your regulator or adviser.

Quick answer: the models and the hosted apps are two different things. The hosted apps from DeepSeek, Qwen and Kimi send what you type to the vendor's servers, so keep personal, client and confidential data out of them unless you have read the current terms for your plan and accepted them. Where the vendor publishes open models, you can run them yourself or through a provider you choose, and then the data stays where you put it. For everyday work on sensitive material, a vendor whose terms you have read and accepted is the simpler path.

The app is not the model

Most confusion comes from treating "DeepSeek" as one thing. It is two.

The first is the hosted app and API. You type into someone else's servers. What happens to your text is set by their privacy policy.

The second is the open model. DeepSeek and Qwen publish model weights that anyone can download. If you run those on your own hardware, or on a cloud account you control, the vendor never sees your prompts.

The Australian Government's February 2025 ban made this distinction too. The Home Affairs direction told government entities to remove DeepSeek products from government-issued devices, and SBS reported it did not apply to the general public. A private business is not bound by it. But the stated reason, a security risk, is the same one you should weigh for your own data.

What each vendor says

DeepSeek. Its privacy policy says it collects, processes and stores personal data in the People's Republic of China. It also lists training its models as a purpose, and says users can opt out of training use by writing to the company. Our review says the same in short form: skip it if you would send sensitive business data to the hosted app or API. The open models are the way to get its capability without that trade. The licence on its R1 weights is MIT, though the distilled versions inherit the licence of the model they were built on.

Qwen. Alibaba's open-weight models are published under Apache 2.0, according to the Qwen project page. Our review notes the catch: a few models are not open, the flagship models are API-only, and hosted use processes data on Alibaba Cloud. Alibaba's own documentation shows that the cloud service has several regions, including Singapore, the US, Germany and mainland China, and says static data stays in the region you call. That is useful. It also means you must choose the region deliberately and confirm the setting before you send anything.

Kimi. Moonshot AI is a Chinese company. Its privacy policy says conversation content is sent to cloud servers and processed by its cloud providers. It does not name a country for storage on the page we read, and it says de-identified conversation data may be used to improve its products, without saying whether that includes training. The policy is published in Chinese. Our review's advice stands: read its own policy before using it for work, and skip it if you have strict rules about where data is processed.

Mistral. Mistral is European, and the comparison is useful because the same questions apply. Its help page says input and output data on some plans is used for training by default, with an opt-out, and that Enterprise starts opted out. The page we read does not say where data is hosted. Our review's catch fits: EU hosting terms need confirming in Mistral's documents or a contract. A European vendor promises nothing until you have it in writing.

None of the four is a verdict. Each is a set of terms that differ by plan and change over time. Read the current page for the plan you would actually use.

What the OAIC asks you to check

The privacy regulator, the OAIC, publishes guidance on commercially available AI products. Four points apply here.

  • Read the terms and settings. Check whether the developer can access what you enter.
  • Check where the servers are. Cloud products may send personal information offshore, which brings cross-border disclosure obligations under the Privacy Act.
  • Watch for training. Some products use customer inputs to improve their models. The guidance says that can be hard to justify for personal information.
  • Do not enter personal information, especially sensitive information, into public chatbots. Minimise what goes in.

That last point is the practical one. Whichever vendor you choose, the cheapest control is not putting client names, health details or financial records into a tool you have not approved. Our AI use policy guide shows how to write that rule in a page.

Self-hosting: what it takes

Running an open model yourself removes the vendor from the data path. It adds work.

You need somewhere to run it. Hugging Face is the common place to find open models and run them on rented hardware. Our review notes the catch: hardware is billed by the hour while it runs, so an always-on GPU adds up. You also need someone who can set it up, keep it patched and judge whether the model is good enough for the task.

For most small teams, that cost outweighs the benefit. Self-hosting makes sense when the data is genuinely sensitive, the task is narrow and repeated, and you have or can hire the skill. A good test is whether you could explain the setup to your insurer or a client in two sentences. If not, you are probably better off with a mainstream assistant on a business plan whose data terms you have read. Our guide to choosing an assistant for the whole team covers that route.

A sensible rule of thumb

  • Public or low-risk material (your own marketing drafts, general research, code with no secrets): any of these tools is a reasonable thing to try, if the terms are acceptable.
  • Internal documents and pricing: use an approved tool on an account whose data terms you have read.
  • Personal, client, health or financial information: not in a hosted app from any vendor you have not vetted. If you need an open model for this, self-host it or have an agency set it up.

Where to start

  1. Compare the tools on the facts. The DeepSeek and Qwen comparison puts the reviews side by side, and you can swap in Kimi or Mistral.
  2. Browse the other options. The AI tools page lists the assistants we review, with the catch for each.
  3. Get help deciding. If you are weighing self-hosting or a data-sensitive use, our Concierge can help match you with the right tools or an implementation agency.

Sources, all read on 7 November 2026: the DeepSeek privacy policy; the Qwen3 project page; Alibaba Cloud Model Studio regions page; the Kimi privacy policy; Mistral's help page on training data; the DeepSeek-R1 model card; SBS, 4 February 2025; OAIC guidance on commercially available AI products.

Share this article

All articles