DeepSeek is a Chinese artificial intelligence company, based in Hangzhou and founded in 2023 by Liang Wenfeng, that publishes large language models under the MIT license. Its R1 reasoning model appeared on January 20, 2025, and its V4 family followed on April 24, 2026, with Pro and Flash versions that accept one million tokens of context. Public use of its hosted chatbot has drawn restrictions from governments, mainly over where user data is stored.
Key takeaways
- DeepSeek releases model weights under the MIT license, so developers can run copies on their own hardware instead of using the company's hosted app.
- The V3 technical report put the final training run at 2.788 million H800 GPU hours, or USD 5.576 million at an assumed USD 2 per hour, and says that figure excludes earlier research and experiments.
- The company's privacy policy, updated February 10, 2026, says user data is stored in the People's Republic of China and may be used to train its models.
- Canada's federal government restricted the app on managed phones in February 2025, and a U.S. Department of Homeland Security order, published in the Federal Register on October 20, 2026, bars DeepSeek products from civilian agencies.
Who runs DeepSeek and how it is funded
Calcalist and an online encyclopedia describe a company based in Hangzhou, started in 2023, and funded by High-Flyer, the quantitative hedge fund that Liang Wenfeng also helped start. Sources disagree on the month: one Israeli business outlet says May 2023, while the online encyclopedia lists July 17, 2023, so this article gives only the year. DeepSeek has not published a funding round the way U.S. labs do, and no outside investment could be confirmed as of October 24, 2026.
The company is a private firm, not a public one, which limits what can be checked. Revenue, user counts, and chip inventories are not in any filing that could be located, and figures that circulate online come from third-party estimates. Those are left out here.
Release timeline from the company's own changelog
DeepSeek's API documentation keeps a dated change log. The table lists the entries that mark a new model generation, with details taken from that log and from the model cards the company posts on Hugging Face.
| Date | Release | Details stated by the company |
|---|---|---|
| December 26, 2024 | DeepSeek-V3 | Mixture-of-experts model, 671 billion parameters in total, 37 billion active per token, pre-trained on 14.8 trillion tokens |
| January 20, 2025 | DeepSeek-R1 | Reasoning model trained from V3-Base, MIT license, 128,000-token context, plus six smaller distilled models |
| August 21, 2025 | DeepSeek-V3.1 | Single model with thinking and non-thinking modes |
| December 1, 2025 | DeepSeek-V3.2 | Upgraded the two standard API model names |
| April 24, 2026 | DeepSeek-V4-Pro and V4-Flash | Preview; Pro has 1.6 trillion parameters (49 billion active), Flash has 284 billion (13 billion active); one million tokens of context; MIT license |
| August 13, 2026 | DeepSeek-V4-Pro general release | Adds three thinking-effort levels and peak and off-peak pricing |
| September 10, 2026 | DeepSeek-V4.1-Flash | Native image understanding; pricing reduced |
| October 17, 2026 | V4.1-Flash-1017 and Search API | Upgrade to the Flash model, pricing unchanged, and a search and fetch interface for developers |
Prices change often and the log does not list per-token figures for the newest models, so the company's pricing page is the reference. The R1 distilled models were built on existing Qwen and Llama base models, which carry their own licenses.
What the training numbers say, and what they leave out
The figure most often quoted comes from the DeepSeek-V3 technical report. Its cost table lists 2.664 million GPU hours for pre-training, 119,000 for context extension, and 5,000 for post-training, a total of 2.788 million H800 hours. At an assumed rental price of USD 2 per GPU hour, the report arrives at USD 5.576 million. The authors state that this covers only the official training run and excludes costs tied to prior research and ablation experiments. It is therefore not the total cost of building the company's models, and it does not compare directly with spending figures other labs report.
R1 received outside review. A paper in Nature, published September 17, 2025 after receipt on February 14, 2025, reports that reinforcement learning alone, without human-written reasoning examples, can improve reasoning on verifiable tasks in mathematics, programming, and STEM fields. The paper also says reasoning patterns from large models can be used to improve smaller ones.
Hosted app or self-hosted weights: a data checklist
Where a DeepSeek model runs decides where the data goes. The points below summarize the company's privacy policy (last updated February 10, 2026) and the licensing terms on its model cards.
- The hosted app and API collect prompts, uploaded files, chat history, device identifiers, and approximate location, according to the policy.
- The policy says the company stores data in the People's Republic of China, and that it uses personal data to train and improve its models.
- Opting out of training is handled by emailing the company; the policy does not describe an in-app switch.
- Open weights can be downloaded and run on an organization's own servers, so prompts need not leave its network. The model still needs its own security and accuracy review.
- Third-party services that host DeepSeek models have their own terms, which are separate from DeepSeek's policy.
Government and institutional responses
In Canada, the Treasury Board Secretariat's chief information officer, Dominic Rochon, wrote to departments on February 6, 2025, describing the step as a precaution over the collection and retention of sensitive personal information, according to Global News. Shared Services Canada had already blocked the app on the mobile devices it manages, and departments outside its support were advised to consider blocking the app and website. Reporting calls this a recommendation to departments rather than a blanket prohibition.
In the United States, security news outlets reported in October 2026 that a Department of Homeland Security order, which most reports date July 24, 2026 although the sources disagree on the signing date, directs civilian agencies to stop buying DeepSeek products and remove them from their systems. It was issued under the Federal Acquisition Supply Chain Security Act of 2018. Reports say agencies can request exceptions and that the order does not apply to private consumers. The Federal Register notice (October 20, 2026, docket DHS-2026-0628) carries the order, but its signing date could not be confirmed from the text reviewed, so the date rests on secondary reporting.
Markets reacted as well. On January 27, 2025, Nvidia shares fell about 17 percent, as CNBC reported, after the R1 chatbot topped the free-app chart in the U.S. App Store. That reaction was about expectations for chip demand, not a measured change in what AI costs to build.
FAQ
Is DeepSeek open source?
Its model weights are released under the MIT license, which permits commercial use and modification. The company's model cards do not describe a release of the full training data, so "open weights" is the more exact term.
Did DeepSeek really train V3 for USD 5.6 million?
The V3 technical report states USD 5.576 million for the final training run at a rental rate of USD 2 per GPU hour. The report says the number leaves out prior research and experiments, so it is not the full development cost.
Where does the DeepSeek app store user data?
The company's privacy policy says it stores data in the People's Republic of China. Organizations with data residency rules should review the policy directly and check applicable law before using the hosted service.
Can DeepSeek models be run without sending data to the company?
Yes, for the open-weight models. A copy run on local or private-cloud hardware does not send prompts to DeepSeek's servers, though hardware needs are large for the biggest models.
Related reading
- Profile of Anthropic and its Claude models
- What the OpenAI research lab publishes
- Open-source AI uses and trade-offs
- The cost of AI for businesses
More in AI Tools and Industry.
Last verified: October 2026
Sources
- DeepSeek, API Docs change log, accessed October 2026
- DeepSeek, Hugging Face model cards for DeepSeek-R1, V4-Pro, and V4-Flash, 2025 and 2026
- DeepSeek-AI, DeepSeek-V3 Technical Report, arXiv:2412.19437, December 2024, revised February 2025
- DeepSeek-AI, DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning, Nature 645, September 17, 2025
- DeepSeek, Privacy Policy, updated February 10, 2026
- Global News, Canada restricts DeepSeek on government devices, February 7, 2025
- BleepingComputer and other security outlets, reports on the DHS order on DeepSeek, October 2026
- Calcalist, profile of DeepSeek founder Liang Wenfeng
- DeepSeek, entries consulted for the founding date discrepancy, October 2026
- CNBC, Nvidia share price reaction, January 27, 2025