DeepSeek is cheap and smart. Then who?

What is DeepSeek

DeepSeek is the name of the large language models (LLMs) and chat service built by DeepSeek (legal name: Hangzhou DeepSeek Artificial Intelligence Co., Ltd.), headquartered in Hangzhou, China. You can use it three ways: free chat (web and mobile app), a low-cost API, and open-weight models whose weights are public.

Which country is it from

It is a Chinese AI. Its privacy policy says the operator is a Hangzhou-based company and that information collected from users is stored on servers in the People's Republic of China. The app is also available on Japan's App Store.

Who founded it

DeepSeek was founded in 2023 by Liang Wenfeng. According to Forbes and other reports, he is a co-founder of the Chinese investment firm High-Flyer. Since R1, the company has been known for publishing its model weights.

Three ways to use it

• Chat (web and app): Talk to it for free at chat.deepseek.com or in the mobile app.
• API: Build DeepSeek's API into your own systems or developer tools. You pay for what you use.
• Open weights: Run the models published on Hugging Face on your own servers or PC. Your data stays inside.

"The DeepSeek shock" refers to the stir after DeepSeek released its reasoning model DeepSeek-R1 in January 2025. Performance close to top US models was available in a free app, and it was reported to have been built with little money and compute. It climbed to the top of the App Store in both Japan and the US.

• Markets: The assumption that powerful AI needs huge chip investment was shaken, and shares of NVIDIA and other chip-related companies fell sharply.
• Technology: The weights were released under the MIT license, so they were built into other companies' services and research, and reasoning models suddenly became accessible.
• Regulation: At the same time, where data is stored and how it answers political topics became issues, and authorities in several countries issued warnings or restrictions (see "Safety" below).

Because the API is cheap and the weights are open, DeepSeek is built into other services, and some people use it indirectly without knowing. It's worth checking which model your work tools run, and where.

1. Open chat.deepseek.com. It works in your browser. On mobile, there is a "DeepSeek" app for iOS and Android.
2. Log in. With an account, your chat history is saved.
3. Start talking. Attach files or images and ask it to "summarize this" or "make this a table".
4. Pick a mode. Toggles for a deeper thinking mode or web search may appear near the input box (labels change over time).
5. Check the training setting. Turning off "Improve the model for everyone" in settings (the label at the time of writing) asks DeepSeek not to use your input to train its models.

As of October 2026, we could not find a paid personal plan for DeepSeek chat.

Tips for using it at work

• Decide what goes in first: Start with public information and general questions, and decide up front that confidential material stays out (see the checklist below).
• State the audience and format first: Saying who it's for and what you need up front, such as "for the leadership meeting, in 3 lines, with sources", cuts down on rework.
• Check the answers: Verify numbers, names, and statements about laws or specs against the original sources before using them.

Turning those answers into who does what by when, and following up, happens outside DeepSeek. If you're doing that by hand every time, see how SIMY connects.

Scroll the table sideways →

API model names: The older deepseek-chat and deepseek-reasoner have been announced for retirement, and users are asked to switch to the new model names. If you have existing code, check the official change log.

• OpenAI compatible: Set the endpoint to https://api.deepseek.com and call it from the OpenAI SDK or existing tools. The official change log also mentions support for the Responses API and Codex.
• Anthropic compatible: With https://api.deepseek.com/anthropic, you can point Anthropic-format tools (such as Claude Code) at DeepSeek.
• How pricing works: Pay-as-you-go in US dollars per million tokens. Input that reuses the same prefix as a previous request (a cache hit) is very cheap, and output costs more than input.
• Peak hours cost double: During weekday peak hours (UTC 01:00–04:00 and 06:00–10:00, excluding Chinese public holidays), rates are twice the off-peak rate. Check how these map to your working hours; moving bulk jobs off-peak keeps costs down.

Rates change often, so check the official DeepSeek pricing page.

Getting started with the API

1. Create an account on the DeepSeek API platform and top up your balance.
2. Create an API key. Keep the key in an environment variable, never directly in code or shared documents.
3. Swap the endpoint and model name. In OpenAI format, set the endpoint URL and a model name such as deepseek-flash.
4. Decide what data to send before scaling up. What you send to the API is also handled under the privacy policy. Check it against your internal rules before moving to production volumes.

How SIMY relates: SIMY does not connect DeepSeek API keys at this time. SIMY's cloud runs use your ChatGPT (Codex) account.

• What runs on a PC: Ollama's deepseek-r1 comes in sizes from 1.5B to 671B, and the 7B version is about 4.7 GB. Run it with a size, such as ollama run deepseek-r1:7b.
• What's cloud-only in Ollama: deepseek-v4.1-flash and deepseek-v4-pro have only cloud tags in Ollama's official library. They run in Ollama's cloud, so they don't stay entirely on your PC.
• Running the latest in-house: V4.1-Flash weights are on Hugging Face and run with vLLM or SGLang, but you need large GPU servers.

If the goal is keeping data inside, only running the weights on your own PC or servers qualifies. Cloud-tagged models don't serve that goal, even if their names appear in Ollama. The Chinese AI (LLM) comparison explains why in "Being in Ollama doesn't mean it runs locally".

For memory requirements and choosing a quantization, see also How to run a local LLM (Qwen).

DeepSeek's model cards don't list supported languages, but in practice you can ask questions and get answers in languages such as Japanese. The app is also available on Japan's App Store.

• Strengths: Answering questions, summarizing long documents, drafting emails and reports, and explaining code.
• Watch out for: Local names, fine points of formal or polite language, and statements about local laws and business customs. Check them before use.
• Small models: The small R1 versions you run locally can produce unnatural text, or show their thinking in English or Chinese.

The surest way to judge quality in your language is a small test with real work documents.

Where data is stored and training

DeepSeek's privacy policy (updated February 10, 2026) says the information it collects is stored on servers in the People's Republic of China. It also states your right to refuse the use of your input for model training, which you can turn off in the app settings. Turning it off doesn't change where data is stored.

Warnings from Japanese authorities

• Japan's Personal Information Protection Commission (February 3, 2025, updated March 5): Published "Information on DeepSeek". It notes that data including personal information collected by DeepSeek is stored on servers in China and is subject to Chinese laws (such as the Personal Information Protection Law, the Data Security Law and the National Intelligence Law).
• Japan's Digital Agency (February 6, 2025): The secretariat of the Digital Society Promotion Council issued a notice to government agencies, "Caution on the business use of generative AI such as DeepSeek". According to reports, it cites DeepSeek as an example of the same risks, says confidential information should not be handled as a rule, and asks agencies that use it to assess the risks first. It is not a blanket ban.

As of October 2026, no law or order in Japan bars individuals from using DeepSeek. If your employer or clients have their own usage rules, follow those.

Actions in other countries (reported)

• Italy: In January 2025, the data protection authority made the app unavailable.
• Australia: In February 2025, it banned use on government devices.
• South Korea: In February 2025, new app downloads were suspended under a measure by its Personal Information Protection Commission.
• Germany: In June 2025, a data protection authority asked Apple and Google to remove the app from their stores.

Most of these concern government devices or the handling of personal data, and the reasons differ by country. Some also point out that its answers lean one way on political topics, so it's safer to use several sources when researching them.

Actions in other countries are summarized from news reports. Check each authority's announcements for the latest. How SIMY handles data is published in Security and our Privacy Policy.

Scroll the table sideways →

• Check company rules first: Some companies' generative AI policies set which services you may use and what you may enter.
• Check client requirements too: Government clients and those in data-sensitive industries may set conditions on the AI provider and where data is stored.
• When in doubt, keep it inside: If you must test with internal information, consider running the weights in-house.

This table is a general guide, not legal advice. For the final call, follow your employer's policies and the responsible department.

Scroll the table sideways →

Which to use: Any AI, DeepSeek included, can produce the answer or draft. If you're tracking owners, deadlines, reports and updates by hand every time after that, that's where SIMY comes in. See the Chinese AI (LLM) comparison for the main Chinese AIs, and the Qwen Guide for Qwen.