Level : Beginner
DeepSeek: what to know before you use it
DeepSeek is a free AI chatbot from a Chinese company that also releases its models as open weights, which makes it an appealing option for reasoning and coding tasks. Before you use it, the key point to remember is that your conversations are stored in China, a fact that has already prompted several European data protection authorities to act.
DeepSeek at a glance
DeepSeek is both a company and a chatbot. The company, Hangzhou DeepSeek Artificial Intelligence Co., Ltd., is registered in China. It became widely known for strong language models released openly with detailed technical reports. Our DeepSeek lab profile covers its two model families.
The chatbot works much like ChatGPT: you type a question at chat.deepseek.com or in the mobile app, and it answers in English or many other languages. To see where it fits among the other assistants, our ranking of the best AI tools and how we built it gives the bigger picture.
Three ideas sum DeepSeek up:
- It is free, with no subscription to unlock.
- The models are open: their weights can be downloaded.
- Your data goes to China when you use the official chatbot or API.
What the free chatbot offers
According to information gathered from the company’s website on October 22, 2026, the DeepSeek chatbot has no paid plan. It searches the web, reads the files you upload and replies in your language. It does not generate images. Agent mode, voice and memory are not specified by the company.
Since an announcement on April 24, 2026, the chatbot has offered two modes, according to our DeepSeek profile:
- Instant, powered by DeepSeek-V4-Flash, for quick answers;
- Expert, powered by DeepSeek-V4-Pro, for reasoning, math and programming.
Expert mode relies on a reasoning model: it works through intermediate steps before giving its final answer, which helps with multi-step problems.
Here is how DeepSeek compares with two widely used assistants, based on our site data:
| Assistant | Free | Paid from | Web search | Agent mode | Voice | Images |
|---|---|---|---|---|---|---|
| DeepSeek | Yes | — | Yes | — | — | No |
| ChatGPT | Limited | $8/mo | Yes | Yes | Yes | Yes |
| Claude | Limited | $20/mo | Yes | Yes | — | — |
DeepSeek models and open weights
What sets DeepSeek apart is its models. They are released as open weights: the parameters that make the model work can be downloaded. According to our lab profile, recent models use the MIT license, a very permissive license that allows modification and commercial use. The full training data and training code, however, are not published.
In practice, a company can install a DeepSeek model on its own servers, or with a hosting provider it chooses, so conversations never pass through DeepSeek’s infrastructure. For an individual, this is much harder: these models need professional-grade hardware.
Here are the main DeepSeek models listed by models.dev as of October 22, 2026, with their release date, context window (how much text they can read at once) and API price:
| Model | Developer | Release | Context (tokens) | Input ($/M tokens) | Output ($/M tokens) | Open weights |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | DeepSeek | October 17, 2026 | 1.05 M | $0.15 | $0.60 | Yes |
| DeepSeek V4 Pro 0813 | DeepSeek | August 13, 2026 | 1.05 M | $0.66 | $1.98 | Yes |
| DeepSeek V4 Flash 0731 | DeepSeek | July 31, 2026 | 1 M | $0.14 | $0.28 | Yes |
| DeepSeek V3.2 | DeepSeek | December 1, 2025 | 128 k | $0.29 | $0.43 | Yes |
The newest, DeepSeek V4.1 Flash, came out on October 17, 2026. According to models.dev, its price per million tokens is $0.15 for input and $0.60 for output, well below the large American models. To put those prices in context, compare them with two well-known closed models:
If you are looking for other open models, our open-source models page sorts them by lab and license.
The sensitive part: your data
This is the question that should drive your decision. DeepSeek’s privacy policy, updated on February 10, 2026, states that personal data is collected, processed and stored in the People’s Republic of China. That includes the content of your conversations and the files you upload.
There is one setting you can change: under Data Controls, you can turn off “Improve the model for everyone” so your chats are no longer used to train the models. That setting does not change where your data is stored. The company does not document a temporary chat mode.
In Europe, this raises questions under the GDPR, the regulation that governs personal data. European law only allows data to leave the EU under specific safeguards. In the US and the UK, the rules are different, but the practical advice is the same: assume anything you type could be stored abroad. To learn what happens to your chats with assistants in general, read our guide on AI and privacy.
What European authorities have decided
Two official decisions directly concern DeepSeek. We read the authorities’ press releases on October 22, 2026.
| Date | Authority | Decision |
|---|---|---|
| January 30, 2025 | Garante per la protezione dei dati personali (Italy) | Immediate limitation on processing Italian users’ data, plus an investigation |
| May 6, 2025 | Berlin Commissioner for Data Protection (Germany) | Asked DeepSeek to remove its apps or make its transfers to China lawful |
| June 27, 2025 | Berlin Commissioner for Data Protection | Reported the apps to Apple and Google as unlawful content in Germany |
In Italy, according to the Garante’s press release of January 30, 2025, the authority found the response from Hangzhou DeepSeek Artificial Intelligence and Beijing DeepSeek Artificial Intelligence insufficient. The companies had said they did not operate in Italy and that European law did not apply to them, which the authority disputed.
In Germany, according to the Berlin Commissioner’s press release of June 27, 2025, DeepSeek had not complied with the May 6 request. The authority therefore used Article 16 of the EU Digital Services Act to report the apps. Its main objection is that data is transferred to China without adequate safeguards, since the EU has no adequacy decision for that country.
Strengths and limits at a glance
| Strengths | Watch-outs |
|---|---|
| Chatbot is completely free | Data stored in China |
| Expert mode for reasoning and code | Regulatory action in Italy and Germany |
| Open-weight models, MIT license | No image generation |
| Low-cost API for developers | No documented temporary chat |
| Web search and file reading | Agent, voice and memory features not specified |
How to use it more carefully
If you decide to try it, a few habits reduce the risks:
- Turn off training in Data Controls before your first conversation.
- Share nothing personal: no full name, address, ID document, medical record or contract.
- Stick to public topics: general knowledge, open-source code, practice exercises.
- Verify important answers against an official source.
- For work use, check with your IT team or your company’s privacy lead first.
Check yourself You want DeepSeek to summarize a coworker's employment contract. Is that a good idea?
No. A contract contains another person’s personal data, and that data would be stored in China according to DeepSeek’s privacy policy. Remove identifying details first, or use an assistant whose data handling meets your company’s rules.
Who is DeepSeek for?
Rather than a one-line verdict, here is a simple guide by profile.
- You are curious and want to try a free reasoning model: DeepSeek works, on topics with no personal data.
- You are a student: useful for practicing math or programming problems, as long as you check each step and do not upload assignments with your name on them.
- You are a developer: the API is cheap and open weights let you host the model yourself. Read each model’s license terms.
- You work with customer, health or HR data: avoid the online chatbot.
- You work for a government body or a regulated business: follow your data protection officer’s guidance.
Alternatives depending on your priority
It all depends on what matters most to you:
- Privacy and EU hosting: Mistral AI’s Le Chat stores data in the European Union by default. Our guide to Le Chat by Mistral covers its plans.
- Free above all: several assistants have a solid free tier, compared in best free AI.
- All-round versatility: ChatGPT and Claude remain the benchmarks, compared in ChatGPT or Claude.
For a head-to-head, our ChatGPT vs DeepSeek comparison puts both on the same criteria. Still unsure? The site quiz points you in the right direction in a few questions:
Frequently asked questions
Is DeepSeek safe to use?
The chatbot itself works like other assistants and can make confident mistakes, so double-check figures, names and important advice. The main concern is privacy: according to its privacy policy, your data is stored in China. Keep personal and work information out of it.
Is DeepSeek banned in the US or the UK?
This article covers the official decisions we were able to read at the source, which come from Italy and Berlin, Germany. Rules can differ by country and by employer. Check your own country's guidance and your workplace policy before installing it.
Can I use DeepSeek models without sending data to China?
Yes, as long as you do not go through the official chatbot or API. The models are released as open weights, so a company can run them on its own servers or with a cloud provider of its choice. This takes serious hardware or a hosting partner, so it is mostly an option for professionals.
Do I have to pay for DeepSeek?
No. As of October 22, 2026, the web and mobile chatbot has no paid plan. Only the developer API is billed, based on usage.