ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›IS DEEPSEEK, KIMI OR QWEN SAFE TO USE?
ARTICLES · GUIDEBEGINNER

Is DeepSeek, Kimi or Qwen Safe to Use?

It depends on how you use it: the app sends data to the company, a download stays on your computer, a US cloud copy stays there. Sourced and dated.

READ9 min
WORDS1,713
SECTIONS7
TYPEGUIDE
CHECKED1 OCT 26

It depends on how you use it. Typing into DeepSeek’s own app or website sends your words to a company that says it stores them in China. Downloading the model and running it on your own computer keeps your words on your computer.

Using a copy hosted by an American cloud, such as Kimi K3 on Amazon Bedrock, keeps them with that cloud company. The model is the same; the route decides who sees your data.

TL;DR — THE SHORT VERSION
  • The plain answer: for everyday questions with nothing private in them, the apps work like any other chatbot. For anything private or work-related, either run the model yourself or use a copy hosted by a cloud company you already trust.
  • The company’s app: DeepSeek’s privacy policy says it collects, processes and stores your data in China.
  • Your own computer: a downloaded model sends nothing back to the company. The risk moves to the file itself, so download it only from the company’s official page.
  • A US cloud copy: Kimi K3 runs on Amazon Bedrock, with an option that keeps requests inside US data centres.
  • Governments: many banned DeepSeek on government devices in early 2025. In July 2026 the White House reportedly considered rules for US companies; this site found none enacted as of 1 October 2026.
  • Allegations are not findings. Anthropic accuses three Chinese labs of copying its model’s abilities; a single report says China’s internet regulator is investigating two of them. Neither is settled.
◈ IN PLAIN TERMS

A model is the AI itself. You can reach it three ways. The app or website runs on the company’s servers, so the company receives what you type. Open weights means the company also publishes the model’s files, so you can download it and run it on your own machine, where nothing is sent anywhere. A cloud host such as Amazon Bedrock runs its own copy of the downloaded model on its servers, under its own rules.

Three ways to use the same model

Who sees your data depends on the route, not on the model’s name.

The three routes, as documented by the companies involved. Read 1 Oct 2026.
RouteWho receives what you typeWhere the data goesExample
The company’s app, website or APIThe companyWherever its privacy policy says: for DeepSeek, ChinaDeepSeek’s app, kimi.com
Download and run it yourselfNobody elseStays on your computer or serverDeepSeek-V4.1-Flash, Kimi K3 or Qwen3.8 from Hugging Face
A copy on a US cloudThe cloud companyThe cloud’s regions you chooseKimi K3 on Amazon Bedrock

DeepSeek, Privacy Policy, last updated 10 Feb 2026, read at source 1 Oct 2026: “To provide you with our services, we directly collect, process and store your Personal Data in People’s Republic of China.” AWS, Kimi K3 model card, Bedrock user guide, read at source 1 Oct 2026. Open weights for each example: DeepSeek, V4.1-Flash card; Moonshot AI, Kimi-K3 card; Qwen, Qwen3.8 on GitHub; read 29 Sep or 1 Oct 2026.

TAKEAWAY

Before asking “is this model safe?”, ask “whose computer is it running on?”

Using the company’s own app

The apps send your words to the company, and their privacy policies say what happens next. DeepSeek’s policy, last updated on 10 February 2026, says it stores personal data in China. It also lists “the right to opt-out of using your Personal Data for training our models or optimizing our technologies.”1

Kimi’s international service is run by NOVASCENT PRIVATE LIMITED, which its policy also calls Moonshot AI. The policy says your information “may be transferred to and stored on servers located outside of your country of residence”, without naming a country, and that it uses your content for “training and optimizing our artificial intelligence (‘AI’) models.”2 This site did not read Qwen’s app privacy policy for this page.

Why the location matters: data held in a country is reachable under that country’s laws. That is the reason many governments gave for their DeepSeek bans (next section). It is a question of jurisdiction, not proof that anyone has looked at your chats.3

TAKEAWAY

Do not type anything into a chatbot app that you would not want stored abroad and used for training. That rule applies to every chatbot, and more strongly where the policy names another country.

What governments have done

The bans so far are about government devices and privacy, and most date from early 2025. An independent summary by MIAI, a research institute, lists them. Among them: the US Navy (24 January 2025), the Pentagon (28 January), Italy’s privacy regulator (30 January), the US House of Representatives (30 January), NASA (31 January), Japan (7 February), South Korea (15 February, when the app was withdrawn from its app stores) and Australia’s government systems (25 February).4 These actions targeted DeepSeek’s app and service, not people running the downloaded model on their own machines.Reasoning, October 2026 — the actions listed concern the app, government devices and data sent to DeepSeek; a downloaded model sends nothing to DeepSeek.

The United States has put Zhipu, the maker of the GLM models, on a trade blacklist. In January 2025 the US Commerce Department added Zhipu AI, the company behind the GLM models sold internationally as Z.ai, and nine related companies to its Entity List. It said they “advance the People’s Republic of China’s military modernization through the development and integration of advanced artificial intelligence research.” Being listed means US firms need a government licence to send them US-regulated technology, and the rule says such licences are presumed to be refused.5 The rule is about exports to those companies. It does not mention people using GLM models.Reasoning, October 2026 — from the rule’s text above; this is not legal advice.

New US rules were discussed in 2026, but none was found in force. Fortune reported on 22 July 2026 that “The White House reportedly considered using executive power to impose conditions on U.S. companies looking to use Chinese AI models”. This site searched on 1 October 2026 and found no such rule enacted.6

TAKEAWAY

If you work for a government or a regulated company, check your employer’s rules first. For everyone else, this site found no law, as of 1 October 2026, that bans using these models.

Allegations you may have read about

Two stories from September 2026 are claims, not findings. Treat them that way.

Anthropic’s distillation accusation. Distillation means training your own model on another model’s answers, to copy its abilities cheaply. On 10 September 2026, TechCrunch reported that Anthropic had detailed “distillation campaigns” from Alibaba, Moonshot AI and DeepSeek, about 200 million exchanges across five campaigns. The article carried no response from the three companies. This site has not read Anthropic’s own report.7 For you as a user, this is a dispute between companies. It says nothing about what happens to your data.Reasoning, October 2026 — the accusation concerns how the models were trained, not how user data is handled.

A reported Chinese investigation. Decrypt reported on 23 September 2026, citing The Information and unnamed sources, that the Cyberspace Administration of China had opened an inquiry into whether DeepSeek and Moonshot AI “secretly routed sensitive user data to Anthropic’s servers”. Neither company nor the regulator confirmed it in that report, and this site found only that one account.8

TAKEAWAY

Allegations can turn out true or false. Base your choice on what the privacy policies and licences say today, and watch for official findings.

Fake downloads, and how to check a file is official

If you download a model, get it from the company’s own page. Anyone can upload files to Hugging Face under a famous name. On 1 October 2026 an account called TypeSafeAI offered “Step-5-Preview-BF16”, said “BF16 checkpoint available now”, and used StepFun’s name and links. StepFun’s own stepfun-ai page had no Step 5 files that day; a news report said its weights were due on 15 October.9 This site did not inspect those files. They may be harmless, but they are not StepFun’s.

Some model files can do harm when opened. Hugging Face’s own documentation warns that the common “pickle” file format can run code on your computer: “There are dangerous arbitrary code execution attacks that can be perpetrated when you load a pickle file.” Its advice is to load models “from users and organizations you trust”. It also scans uploads, but says the scan is “not 100% foolproof”.10

Four checks before you download:

  1. Start from the company. Open the model’s announcement on the company’s own site or GitHub and follow its link to the files. Do not start from a search result.
  2. Check the account name. Official pages include deepseek-ai, Qwen, moonshotai, zai-org, MiniMaxAI, tencent, XiaomiMiMo and stepfun-ai on Hugging Face.
  3. Check the date. If the company says the weights are not out yet, a download that says otherwise is not from the company.
  4. Prefer safetensors files. Hugging Face created the safetensors format as a safer way to store weights than pickle.

Account names from each company’s own Hugging Face organisation, as linked on AI labs outside the US and the Chinese AI models table, read 29 Sep – 1 Oct 2026. Hugging Face, Pickle Scanning, read 1 Oct 2026: “we are experimenting with a new, simple serialization format for weights called safetensors”.

TAKEAWAY

The account name matters more than the model name. If you did not arrive from the company’s own link, go back and do.

What to do, in order

  1. Nothing private? The app is fine for casual use; read its privacy policy once.
  2. Private or work data? Use a downloaded copy on your own machine, or a cloud copy with a region you choose, such as Kimi K3 on Amazon Bedrock’s US-only profile.
  3. Working for a government or regulated firm? Ask first; many public bodies banned DeepSeek’s app in 2025.
  4. Downloading? Only from the company’s official page, and preferably safetensors files.

Reasoning, October 2026 — a summary of the sourced sections above. See also what happens to what you type.

What this page could not verify

  • Anthropic’s own report. Only TechCrunch’s account of it was read.
  • The Chinese investigation. One outlet, citing unnamed sources.
  • Qwen’s and Z.ai’s app privacy terms in detail. Not read for this page, beyond Z.ai saying its international service is run from Singapore.
  • Whether any downloaded model contains hidden behaviour. Nobody outside the companies can fully check how a model was trained.
  • The exact date the Entity List rule took effect. The public-inspection copy read here says it takes effect on publication.

Related on this site: Is ChatGPT Safe to Use?

SOURCES
  1. 1DeepSeek, Privacy Policy, read at source 1 Oct 2026.↑
  2. 2Kimi, Privacy Policy, last updated 13 Aug 2026, read at source 1 Oct 2026.↑
  3. 3MIAI, DeepSeek one year later, 28 Jan 2026, read at source 29 Sep 2026“data inputted into the model will be sent to China and thus is subject to Chinese law”. ↑
  4. 4MIAI, DeepSeek one year later: regulatory storm, global surge, 28 Jan 2026, read at source 1 Oct 2026.↑
  5. 5US Bureau of Industry and Security, Addition of Entities to and Revision of Entry on the Entity List, Federal Register document 2025-00704, read at source 1 Oct 2026“A license is required for all items subject to the EAR, with a license review policy of a presumption of denial.” ↑
  6. 6Fortune, As Washington panics about Chinese AI, Jensen Huang says open-source models like Kimi are ‘excellent’, 22 Jul 2026, read at source 1 Oct 2026. Web search for an enacted rule, 1 Oct 2026: none found.↑
  7. 7TechCrunch, Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek, 10 Sep 2026, read at source 1 Oct 2026, quoting Anthropic“persistent distillation attacks by China-based AI companies”. ↑
  8. 8Decrypt via Yahoo News, China Probes DeepSeek and Moonshot Over Alleged Data Leaks to Anthropic’s Claude, 23 Sep 2026, read at source 1 Oct 2026.↑
  9. 9Hugging Face, TypeSafeAI/Step-5-Preview-BF16 and the stepfun-ai listing, both read at source 1 Oct 2026; MarkTechPost, report on Step 5 Preview, 20 Sep 2026, read at source 29 Sep 2026.↑
  10. 10Hugging Face, Pickle Scanning, Hub documentation, read at source 1 Oct 2026“We suggest loading models from users and organizations you trust”. ↑

Companies: DeepSeek privacy policy and V4.1-Flash card · Kimi privacy policy and Kimi-K3 card · Qwen Qwen3.8 GitHub · AWS Bedrock Kimi K3 card · Hugging Face Pickle Scanning, TypeSafeAI copy and stepfun-ai listing.

Government and press: US Bureau of Industry and Security, Entity List rule 2025-00704 · MIAI, DeepSeek one year later · Fortune, 22 Jul 2026 · TechCrunch, 10 Sep 2026 · Decrypt via Yahoo News, 23 Sep 2026 · MarkTechPost, 20 Sep 2026. Read at source 29 Sep or 1 Oct 2026, as marked beside each claim. Allegations are reported as allegations.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE