The short answer
DeepSeek is a free AI chat assistant from Hangzhou DeepSeek Artificial Intelligence Co., Ltd. in China. Since 17 October 2026 the web and app run DeepSeek-V4.1-Flash-1017, and Expert Mode switches to the larger DeepSeek-V4-Pro, which accepts up to 1 million tokens of context. It is strong at reasoning and code, but its privacy policy says user data is stored in China, so keep confidential work out of it.
| Made by | Hangzhou DeepSeek Artificial Intelligence Co., Ltd., China |
|---|---|
| Free plan | Yes, the web chat and apps are free |
| Models | DeepSeek-V4.1-Flash-1017 by default, DeepSeek-V4-Pro in Expert Mode |
| Best at | Free reasoning, maths and code; long documents; open weights for self-hosting |
| Shows sources | Has web search; how it shows sources is not documented |
| Team plans | None for the chat; API and self-hosted open weights instead |
| Your data | Stored in China under PRC law; inputs may improve the service unless you turn off Improve the model for everyone |
What DeepSeek is in November 2026
DeepSeek is the AI lab run by Hangzhou DeepSeek Artificial Intelligence Co., Ltd. in China, and its chat assistant at chat.deepseek.com and in the iOS and Android apps is free to use. The App Store lists the app as free, and DeepSeek publishes no paid consumer subscription; the only thing it sells on its own site is API access for developers.
The models changed a lot this year. DeepSeek previewed its V4 family on 24 April 2026, made DeepSeek-V4-Pro generally available on the web, in the app and in the API on 13 August, released V4.1 Flash with native image understanding on 10 September, and moved the web and app to DeepSeek-V4.1-Flash-1017 on 17 October. In practice that means two things for a chat user:
- V4.1 Flash is what you get with Expert Mode off. It is fast, it can read images, and DeepSeek says the October update gives shorter outputs and faster task completion.
- V4-Pro is the bigger model, reached by switching on Expert Mode. It has 1.6 trillion parameters, 49 billion of them active for each token, and it handles text only.
DeepSeek's model card gives V4-Pro up to 1 million tokens of context, which is enough for several long reports at once; DeepSeek's V4.1 Flash release note gives no figure. Both think step by step by default when used through the API, with low, high and max effort levels.
The second thing that makes DeepSeek unusual is that it gives the models away. The V4-Pro weights are on Hugging Face under the MIT licence, which lets any company download them and run them on its own servers or with a cloud provider of its choice. That matters for privacy, as the section below explains.
What DeepSeek is genuinely good at
Reasoning through a problem for free
For maths, logic, pricing scenarios or "what happens to margin if we change three variables", DeepSeek with Expert Mode on is a strong free option. Ask it to show its working and check the steps, which is easy because it lays them out in order.
Code and technical questions
DeepSeek pitched its August V4-Pro release at agent work, with one-click setup for Codex, and DeepSeek has built its API to work with OpenAI and Anthropic formats and with tools such as Codex and Claude Code. For a quick script, a spreadsheet formula or an explanation of an error message, the free chat is often all you need.
Long documents that contain nothing sensitive
The 1-million-token context means you can paste a long public report, a published standard or an open-source codebase and ask questions across all of it. Keep this to material that is already public, for the reasons below.
Running a capable model inside your own walls
For technical teams, the open weights are the real story. A company that self-hosts DeepSeek-V4-Pro gets a frontier-class model where no prompt leaves its own infrastructure. That is a very different privacy picture from using the free app, and it is the main reason serious organisations look at DeepSeek at all.
Where DeepSeek falls short
- Data residency. Everything you type in the app or web chat is stored in China. This is the main reason most companies should not use the consumer app for work.
- No work plan. There is no team or enterprise version of the chat with admin controls, single sign-on or a contractual training exclusion. ChatGPT and Mistral's Vibe, among others, offer one.
- Few connections to your tools. The chat does not connect to your email, calendar or drive the way ChatGPT, Perplexity or Mistral's Vibe can, so it stays a place you paste things into.
- V4-Pro cannot see images. If you need to work from a screenshot or a scanned page, stay on the default Flash model.
- Restrictions are spreading in government. On 19 October 2026, the US Department of Homeland Security made public an order, dated 24 July and published in the Federal Register the next day, telling federal civilian agencies to remove DeepSeek technology from their systems and stop buying it, citing the risk of sensitive data going to China. China's foreign ministry objected. It does not apply to private companies, but it tells you how security teams view the app.
DeepSeek data residency and privacy, plainly
This is the section to read before you paste anything into DeepSeek. All of it comes from DeepSeek's own privacy policy (last updated 10 February 2026) and Terms of Use (last updated 27 March 2026).
- Where your data goes. The privacy policy says DeepSeek collects, processes and stores personal data in the People's Republic of China.
- Which law applies. The Terms of Use are governed by the laws of mainland China.
- Training. The privacy policy lists training and improving its technology as a purpose. The Terms say inputs and outputs may be used to a minimal extent, after encryption and de-identification, to improve the service.
- How to opt out. The Terms say you can opt out by turning off the setting called Improve the model for everyone in the app's settings. DeepSeek does not publish the exact menu path, so look for it in the settings or data section of the app. You can also email a training opt-out request to privacy@deepseek.com, which the privacy policy names as the contact for privacy requests.
- How long it keeps data. Account data, inputs included, is kept for as long as you have an account, and DeepSeek may keep it longer if you break its terms.
- What the app collects. The App Store privacy label lists data linked to you including email address, phone number, coarse location, photos or videos, audio, search history, user content and device ID.
- Age. The Terms say the service is mainly for adults and minors need a guardian's consent.
Switching off the training toggle is worth doing, but be clear about what it does not change: your chats are still stored in China. The honest rule is simple. Use the free app for public information, learning and personal side projects. Do not put in customer data, contracts, financials, health information, passwords or anything your employer has not approved. If your organisation wants DeepSeek's models for real work, the safer route is to run the open weights on infrastructure you control, so your data never reaches DeepSeek at all.
Using DeepSeek's open weights instead of the app
If your team likes DeepSeek's answers but cannot accept data stored in China, the open weights are the way through. The model card on Hugging Face says the repository and weights are licensed under MIT, which allows commercial use. Your IT or engineering team can run the model on its own servers, or on a cloud account the company already controls, and point internal tools at it. Prompts and documents then stay wherever that infrastructure lives, and DeepSeek's privacy policy no longer applies to them, because DeepSeek never receives them.
Two cautions. First, V4-Pro is very large, so running it yourself needs serious hardware; the smaller Flash models are cheaper to host. Second, open weights do not remove every risk. Your security team should still review the model, test its answers on your own tasks and set the same rules for checking output that apply to any AI tool.
How to get better results from DeepSeek
- Choose the model on purpose. Turn on Expert Mode for hard reasoning, analysis and code. Leave it off for quick answers and anything with an image.
- Ask it to show its working. "Work through this step by step, then give the answer in one line" makes mistakes easy to spot.
- Strip identifying details first. Replace names, company names and numbers with placeholders such as [client] before you paste. The answer is usually just as useful.
- Use web search for anything recent. Without it, the model answers from what it learned in training, which can be out of date.
- Give it constraints. Say the length, the format and the reader, just as you would with any assistant. DeepSeek tends to be brief, so ask for detail where you want it.
- Check facts elsewhere. Like every chatbot, it can sound sure when it is wrong. Confirm names, numbers and dates against a source, or run the question through Perplexity for citations.
Our free lessons at /learn?tool=deepseek cover these habits with short exercises.
Who DeepSeek suits, and who should skip it
Good fit: students, hobbyists and professionals who want a strong reasoning and coding model for free on public material, and developers who want a capable model with open weights or a low-cost API.
Not a good fit: anyone in a regulated industry, anyone handling client or employee data, public-sector staff, and any team that needs admin controls, an audit trail or a contract that excludes training. For those, a work plan from ChatGPT, Perplexity or Mistral's Vibe, which keeps data in the EU by default, is a better choice.
For side-by-side detail, read ChatGPT vs DeepSeek and Claude vs DeepSeek, or browse every tool in our comparison tool.
Prompts to copy
Work through a pricing scenario
Turn on Expert Mode first. We sell [product] at [price] with [cost] unit cost and [volume] units a month. Work through what happens to monthly profit if we cut the price by 10%, 15% and 20% and volume rises by [x]%. Show each step, then give a one-line recommendation.
You get: A step-by-step scenario you can check line by line
Explain an error message
I got this error in [language or tool]: [paste error, with any names or keys removed]. Explain in plain English what it means, the three most likely causes and how to fix each one.
You get: A plain-English diagnosis with fixes
Write a spreadsheet formula
In [Excel or Google Sheets], column A has [data], column B has [data]. Write a formula that [goal]. Explain how it works and give one example of the result.
You get: A working formula and an explanation
Question a long public report
Here is a public report: [paste or upload]. Give me the five findings that matter most for [audience], with the page or section each comes from, and list any claims the report makes without data.
You get: The key findings with locations and weak spots
Check your own reasoning
Here is my argument for [decision]: [paste]. Find the weakest assumption, the strongest counter-argument and one piece of evidence that would change my mind. Be direct.
You get: An honest stress test of your argument
Learn a concept fast
Teach me [concept] as if I am a smart professional with no background in it. Use one everyday analogy, one worked example and three quiz questions at the end. Wait for my answers before you reveal the solutions.
You get: A short lesson with a self-test
Anonymise before you paste
Before I share a document, rewrite this text so it contains no names, company names, email addresses, phone numbers or exact figures. Replace each with a placeholder like [CLIENT_1] and give me the list of replacements at the end. [paste text]
You get: A cleaned version safe to use with any AI tool
Draft a technical plan
I want to build [small tool or script] that [goal] using [language]. Give me the file structure, the main functions with one-line descriptions, and the first function written in full with comments.
You get: A build plan and a first working piece of code