What Is DeepSeek? Models, Access, Pricing and Privacy

DeepSeek is a Chinese artificial intelligence company based in Hangzhou, and also the name of its chatbot. You can use it through a website, mobile apps or a developer API. Because it publishes its model weights, you can also run its models on your own hardware. It drew global attention in early 2025, and also regulatory scrutiny over where user data is stored.

A note on timing: this guide is based mainly on developments up to mid-2025. Model versions, prices and government actions change often, so check DeepSeek's official documentation and relevant regulators for the latest status.

Who owns DeepSeek?

DeepSeek was founded in July 2023 by Liang Wenfeng, a co-founder of the quantitative hedge fund High-Flyer. The fund bankrolls the company, and as of 2025 it was not known to have raised outside venture capital.

The lab presents itself as research-driven and releases papers and weights publicly. Claims about its staff size and how many Nvidia GPUs it controls are disputed, and analysts such as SemiAnalysis have challenged the narrative that it built frontier models on a shoestring.

What are DeepSeek V3 and R1?

After early language and coding models in late 2023, DeepSeek released V2 in May 2024, which triggered a price war in China. DeepSeek-V3 followed on 26 December 2024. It is a Mixture-of-Experts model with 671 billion total parameters, of which about 37 billion are active per token, and a 128K context window.

The V3 paper reported roughly 2.788 million H800 GPU-hours, or about US$5.6 million, for training. That number covers only the final training run, not research, earlier experiments or hardware. It is often misreported as the model's total cost.

DeepSeek-R1, released on 20 January 2025, is a reasoning model reported to perform comparably to OpenAI's o1 on math and coding benchmarks. Smaller distilled versions based on Qwen and Llama, from 1.5B to 70B parameters, came with it. On 27 January the app topped the US App Store and Nvidia lost about $590 billion in market value in a single day. Updates such as V3-0324 and R1-0528 followed, and newer releases may exist, so check api-docs.deepseek.com/news for the current model.

How do you use DeepSeek?

The simplest route is the free web chat at chat.deepseek.com, where you can sign up with email or a Google account. A "DeepThink (R1)" button switches on reasoning mode, and a "Search" toggle lets it pull from the web. Apps are available for iOS and Android.

Developers can use an OpenAI-compatible API at platform.deepseek.com, which offers deepseek-chat for general tasks and deepseek-reasoner for step-by-step reasoning. The service has suffered "server busy" overloads, and API top-ups were paused for a period in February 2025. The app has also been pulled from stores in some countries.

Is DeepSeek free? How much does the API cost?

The consumer chat is free, and as of 2025 there was no paid subscription tier. It is worth confirming whether that has changed.

The API charges per token. Early-2025 rates per million tokens were about $0.27 input ($0.07 on a cache hit) and $1.10 output for deepseek-chat, and about $0.55 input ($0.14 cached) and $2.19 output for deepseek-reasoner. Off-peak discounts of 50 to 75% have also been offered.

At the time, R1 was roughly 20 to 30 times cheaper than OpenAI o1's list price. Later 2025 releases reportedly unified and lowered prices, so check api-docs.deepseek.com/quick_start/pricing before budgeting.

Is DeepSeek open source?

DeepSeek releases open weights on Hugging Face and GitHub. R1's code and weights use the MIT license, which permits commercial use and distillation. V3's weights initially had a custom model license, with some later releases moving to MIT, so check the license of each model you use.

Open weights are not the same as fully open source, because the training data and complete training code are not published. You can run the models locally with tools such as Ollama, LM Studio or vLLM. The full 671B model needs multi-GPU servers, so most individuals use distilled versions. Cloud providers including Azure AI Foundry and AWS Bedrock also host DeepSeek models.

Is DeepSeek safe to use?

DeepSeek's privacy policy says chats, uploaded files, device information and even keystroke patterns are stored on servers in China. Chinese laws such as the 2017 National Intelligence Law can require companies to cooperate with the state.

Italy's data protection authority ordered a block in January 2025. Ireland, France, the Netherlands and Belgium opened GDPR inquiries. South Korea suspended new downloads in February 2025 and restored them after fixes around April. In June 2025, Berlin's data commissioner asked Apple and Google to remove the app.

Australia, Taiwan and South Korean ministries banned it on government devices. In the US, NASA, the Navy, the Pentagon, the House and states including New York, Texas and Virginia did the same, and a federal bill was introduced. An April 2025 report by the House Select Committee on the CCP called DeepSeek a national security threat.

Researchers also flagged problems. Wiz found an exposed database containing chat logs, NowSecure reported unencrypted data from the iOS app, and Cisco said its jailbreak tests succeeded every time. The hosted chatbot deflects questions on topics such as Tiananmen and Taiwan, and some of that censorship remains in the open weights. OpenAI and Microsoft said they were examining whether DeepSeek used OpenAI outputs, which remains an unproven allegation.

A cautious approach: on the app or API, assume your data is stored in China and avoid entering personal, confidential or work information. Running the open weights locally or through EU or US hosts keeps your data off DeepSeek's servers.

Can businesses use DeepSeek under data protection rules?

Companies sending personal data to the China-hosted API should review cross-border transfer rules in their jurisdiction, such as GDPR in Europe or Türkiye's updated KVKK framework. Self-hosting the open weights is the usual way to avoid that transfer entirely.