Aiodic

Back to the table

62 Model family Dk DeepSeek Model Families

DeepSeek

DeepSeek's model family.

What it is

DeepSeek is a Chinese artificial intelligence company, and also the name of the family of AI models it makes. Its models can chat, write, answer questions, write code, and work through hard maths and logic problems. You can use them in DeepSeek's own free app and website, through its paid API (the doorway developers use to build a model into their own software), or by downloading the models and running them yourself.

That last option is what makes DeepSeek unusual among the best-known model makers. Most of its major models are released as "open weights": the trained model files are published for anyone to download, study, adapt, and run on their own computers. OpenAI, Anthropic, and Google mostly keep their leading models on their own servers. DeepSeek, like Meta with Llama and Alibaba with Qwen, gives the files away.

Where it came from

DeepSeek was founded in Hangzhou, in eastern China, on 17 July 2023 by Liang Wenfeng, who is also its chief executive. It grew out of High-Flyer, a quantitative hedge fund Liang co-founded in 2015. A quantitative fund uses mathematical models and computers, rather than human stock-pickers, to decide what to buy and sell, so High-Flyer had already built up large banks of powerful computer chips and staff who knew how to use them. That hardware and know-how became DeepSeek's starting point, and High-Flyer remains its owner and funder rather than outside venture investors.

The generations, in brief

You don't need to memorise these. The dates come from DeepSeek's own announcements, and the names have followed a pattern: "V" models are general chat models, and "R" models are reasoning models, which work through a problem step by step before they answer (see Extended Thinking).

The long-rumoured "R2" never appeared under that name. Since V3.1, reasoning has been a mode inside the main models instead of a separate product.

Open weights and the licence

"Open weights" is not quite the same as "open source". Open source traditionally means you get everything needed to rebuild a program from scratch. DeepSeek publishes the finished model files and technical reports, but not the full training data or all of the training code, so nobody outside the company can exactly reproduce how the models were made.

What it does give is generous. When R1 came out, DeepSeek released the code and weights under the MIT licence, one of the most permissive licences there is, and said openly that people could distil (train smaller models on its outputs) and use it commercially. The original V3 had used DeepSeek's own custom licence, which included a list of forbidden uses; from V3-0324 in March 2025 the company switched to MIT. The V4 model pages state the MIT licence too. Licence terms can change between versions, so read the licence on each model's own page before building anything on it.

One detail trips people up. Alongside R1, DeepSeek released six smaller "distilled" models, from 1.5 billion to 70 billion parameters. These are not shrunken copies of R1. They are other companies' open models, Qwen and Llama, retrained on examples of R1's reasoning, and they keep their original licences: Apache 2.0 for the Qwen-based ones and Meta's Llama licence for the Llama-based ones. They are the versions most people can actually run on a home computer; the Parameters page explains what size of model fits on what kind of machine.

Why it matters

It changed what people thought a model had to cost. DeepSeek showed that a small company, working under limits on which chips it could buy, could build models that it said came close to the best American ones, and then charge far less for them or give them away. How much it really spent is still argued over, as the Mixture of Experts page explains. But the price pressure it created was real, and it pushed other companies to cut prices and release more open models.

It made open models a serious choice. Before R1, the strongest reasoning models were only available on their makers' servers. Afterwards, anyone with enough hardware could run one privately. The US National Institute of Standards and Technology reported in September 2025 that downloads of DeepSeek models on model-sharing sites had grown nearly 1,000 percent since January 2025.

Where your data goes depends on how you use it. This is the most practical point on this page. DeepSeek's own app and website are run by a Chinese company, and its privacy policy says it collects, processes, and stores users' personal data in the People's Republic of China. Running the open weights on your own computer is completely different: nothing you type leaves your machine. Using a third-party service that hosts DeepSeek models is different again, because then that company's privacy policy and server locations apply. "Using DeepSeek" can mean any of the three.

It sits in the middle of a political contest. DeepSeek has become a symbol in the rivalry between the United States and China over AI and computer chips. That brings government bans, security warnings, and accusations, some tested and some not. Knowing which claims come from whom helps you make your own judgement rather than borrowing someone else's.

Real-world examples

The week DeepSeek topped the App Store (January 2025)

DeepSeek launched its free chatbot app for phones on 10 January 2025. Ten days later, on 20 January, it released R1 and made it available in the app, the website, and the API, with the weights free to download under the MIT licence.

By 27 January 2025, MacRumors reported that DeepSeek had reached number one on Apple's free App Store chart in the United States and several other countries, ahead of ChatGPT. The same day, investors questioned how many expensive chips AI companies would really need, and Nvidia's shares fell by about a sixth in a single day. The Inference and Mixture of Experts pages tell that part of the story.

Within days came the arguments that followed DeepSeek from then on. OpenAI said it was looking into whether DeepSeek had "inappropriately distilled" its models, a claim covered on the Synthetic Data page. And on 29 January 2025 the cloud security company Wiz published research showing it had found a DeepSeek database open to the internet with no password, holding more than a million lines of logs, including users' chat histories in plain text and secret keys. Wiz said it reported the problem and DeepSeek promptly secured it. Leaving a database open is a basic mistake that many fast-growing companies make, but it set the tone for how regulators would look at the app.

Regulators and governments respond (2025)

Europe moved first. Italy's data protection authority, the Garante, had asked DeepSeek what personal data it collected, where it came from, and whether it was stored in China. The companies behind the app replied that they did not operate in Italy and that European law did not apply to them. The Garante called the answers insufficient and on 30 January 2025 ordered an immediate block on DeepSeek processing Italian users' data, which took the app out of Italian app stores.

Then came bans on government devices, which are narrower than a national ban: they stop officials from using an app on work phones and computers but leave the public free to use it. Taiwan's Ministry of Digital Affairs banned it across government agencies, public schools, and state-owned companies at the end of January 2025. Australia ordered it removed from government systems on 4 February 2025, and South Korea's ministries followed with their own restrictions. In the United States, TechCrunch reported that the Navy, NASA, the House of Representatives' administrative office, and the state of Texas had restricted or banned it by early February.

South Korea went further. Its data protection regulator, the Personal Information Protection Commission, suspended new downloads of the app from 15 February 2025, after finding that user data had been sent to ByteDance, TikTok's parent company, without proper consent. DeepSeek acknowledged it had not taken Korean privacy law fully into account and appointed a local representative. After it changed its privacy policy, the app returned to Korean app stores in late April 2025.

On 27 June 2025, Berlin's data protection commissioner, Meike Kamp, formally reported the app to Apple and Google as illegal content in Germany, saying DeepSeek transferred German users' data to China unlawfully under EU rules and had ignored a request to fix this.

Notice what these actions were about. They concerned the hosted app and where its data went, not the downloadable models. None of them stopped anyone running the open weights on their own computer.

Censorship tests and a US government evaluation (2025)

Journalists noticed quickly that DeepSeek's app would not discuss some subjects sensitive to the Chinese government, such as the 1989 Tiananmen Square crackdown or Taiwan's status. China's 2023 rules for generative AI services require their content to reflect the state's "core socialist values", so this was not a surprise, but independent testers set out to measure it.

On 28 January 2025 the AI testing company Promptfoo published results from 1,360 prompts on topics it considered sensitive in China. It reported that R1, accessed through a third-party service rather than DeepSeek's own app, refused or gave a scripted answer to about 85 percent of them, and that the censorship was "brittle" and easy to get around, for example by framing a question as fiction. In May 2025 a team of academic researchers published a paper titled "R1dacted" showing that R1 still refused certain politically sensitive prompts when run locally, so at least some of the behaviour is trained into the model and doesn't vanish when you download it.

On 30 September 2025 the US government's Center for AI Standards and Innovation, part of NIST, published its own evaluation of DeepSeek R1, R1-0528, and V3.1 against American models. It reported that DeepSeek's models echoed four times as many inaccurate and misleading Chinese Communist Party narratives as the US models it compared them with, were far easier to "jailbreak" into answering overtly harmful requests, and, as agents, were more likely to follow malicious instructions hidden in their inputs. It also found the best US model outperformed the best DeepSeek model on almost every test it ran. This was an official US evaluation carried out during a period of open technology rivalry with China, so read it as one careful source among several rather than the last word.

Chips and export controls (2022 to 2026)

To understand DeepSeek you need a little background on chips. Since October 2022 the US has restricted exports of the most powerful AI chips to China. Nvidia designed a slightly slower chip, the H800, to sell there within the rules, and on 17 October 2023 the US banned that too. DeepSeek's reports say V3 and R1 were trained on H800s. Nvidia told Reuters in 2025 that its review indicated DeepSeek had used "lawfully acquired H800 products, not H100".

The accusations have gone further. In June 2025 a senior US State Department official, speaking anonymously to Reuters, said DeepSeek supported China's military and intelligence operations and had tried to use shell companies in Southeast Asia to get around export controls. DeepSeek did not respond to Reuters' questions. These were claims by an unnamed official, not findings from a court.

Meanwhile China has pushed its companies towards home-grown chips. In August 2025 the Financial Times reported, citing unnamed sources, that DeepSeek had struggled to train a new model on Huawei's Ascend chips and fell back to Nvidia chips for training. When V4 arrived in April 2026, Huawei said its Ascend systems would fully support it, Reuters reported. Exactly which chips trained which model is not something outsiders can verify, so treat confident claims either way with care.

An everyday case: the novelist and the manuscript

Picture Tom, who is writing a novel and wants help tightening a chapter. He has heard DeepSeek is free and good at writing, but the chapter is unpublished and includes details from his own family's life. He reads DeepSeek's privacy policy and sees that data from the app is stored in China, which he isn't comfortable with for this. So he installs a free app on his laptop and downloads a small distilled DeepSeek model instead. It is slower and less polished than the big hosted version, but his chapter never leaves his computer. For quick questions that contain nothing private, he still uses whatever chatbot is handy. He hasn't decided DeepSeek is good or bad. He has decided that the way he uses it matters more than the name.

Try it yourself

This exercise compares a hosted chatbot with a DeepSeek model running on your own computer, using a harmless question. The local part needs a reasonably recent laptop or desktop with at least 8 GB of memory and about 6 GB of free disk space. App names, model sizes, and download sizes were correct at the time of writing, but they change, so check the current details on each site.

A privacy note before you start. If you use DeepSeek's own app or website, its privacy policy says your data is stored in China. Some workplaces and governments forbid it on work devices. Don't type anything personal, confidential, or work-related into any hosted chatbot during this exercise, and check your employer's rules first.

  1. Pick your hosted chatbot. Use any chatbot you already have, or DeepSeek's own website at chat.deepseek.com if you're happy with its privacy terms. Another option is a third-party service that hosts DeepSeek models; if you choose one, check where it says your data is processed.
  2. Ask it a harmless question and copy the answer into a document: > I have 3 litres of water in a jug and pour out a third of it, then add 1.5 litres. How much is in the jug? Show your working, then explain in two sentences how you checked your answer.
  3. Install Ollama, a free program that runs open-weight models on your own computer. Download it only from its official site, ollama.com, and install it like any other program.
  4. Download and run a small distilled DeepSeek model. Open your computer's terminal (on Windows, search for "Terminal" or "PowerShell"; on a Mac, open "Terminal") and type: > ollama run deepseek-r1:8b At the time of writing this downloads about 5 GB. If your computer struggles, try deepseek-r1:1.5b instead, which is about 1 GB but much less capable. Once it has downloaded, you can switch off your internet connection: everything now happens on your own machine.
  5. Paste the same question and wait. You'll see the model's reasoning appear first, often between <think> tags, before its final answer. (The answer is 3.5 litres.)
  6. Compare the two. Which was faster? Which explained itself more clearly? Did the local model get it right? Small distilled models make more mistakes than big hosted ones; that trade-off is the price of privacy and control.
  7. Optional: see the political filtering for yourself. Ask both models: "What happened in Tiananmen Square in Beijing in June 1989?" Compare the answers with an encyclopaedia. Note that this small model is built on a Qwen base, so it may behave differently from the full R1 that researchers tested.
  8. Clean up if you like. Type /bye to leave the model, and ollama rm deepseek-r1:8b to delete it and free the disk space.

Common mistakes and misunderstandings

Thinking the app and the model are the same thing. The DeepSeek app is a service run by a Chinese company, with data stored in China under its privacy policy. The DeepSeek models are files you can run anywhere. Bans and privacy warnings have mostly been about the app. Before you decide whether "DeepSeek is safe", ask which of the two you mean.

Assuming running it locally removes every concern. Running the weights yourself solves the data question, because nothing leaves your machine. It doesn't remove the political slant that testers found trained into the model, and US testers also reported that DeepSeek models are easier to trick into harmful answers than the American models they compared. If you build something on them, add your own guardrails.

Believing "trained for 6 million dollars". DeepSeek's figure covered only the final training run of V3, and its later figure of about 294,000 dollars covered only the reasoning stage of R1 on top of V3. Neither includes research, failed experiments, staff, or the chips themselves. The models were efficient, but not that cheap.

Treating every claim about DeepSeek as settled. DeepSeek's benchmark claims are its own. The accusations about distillation, military links, and chip smuggling come from rival companies and US officials, and have not been tested in court. Governments on both sides have reasons to tell a particular story. Credit each claim to whoever made it.

Calling it open source. DeepSeek's models are open weights under a permissive licence, which is very useful, but the training data is not published. And the small distilled models carry Qwen's or Meta's licences, not DeepSeek's.

Expecting it to stay the same. DeepSeek renames, merges, and retires models quickly: at the time of writing, older API model names already redirect to newer models. Check the company's current model list and pricing page before relying on a particular version.

Related elements