DeepSeek statistics: 399 million model downloads and 129 million users in China

DeepSeek's open models have been downloaded 399 million times

The 399 million is our own sum of the download counter on every model in DeepSeek's Hugging Face account, read on 12 October 2026. Its Android app has over 100 million installs, per DeepSeek's own Google Play listing. Every row in these DeepSeek statistics names its source and date. DeepSeek doesn't publish user numbers, so each user figure is someone else's estimate, and we say whose.

R1 is still DeepSeek's biggest hit on Hugging Face. If you count every R1 release and every distilled model built from it, the R1 line has about 176 million downloads, or 44% of the total. The six distills from January 2025, which DeepSeek built on Qwen and Llama at 1.5 to 70 billion parameters, have about 101 million on their own. That's twice the 49.7 million of the original R1 repository. A 7B distill runs on a laptop. The full R1, which is built on the 671 billion parameter V3, doesn't.

The newer models lead from month to month. In the 30 days up to 12 October 2026, the V4 line took 8.4 million downloads and V3 took 8.0 million, while all of the R1 models together took 4.3 million. Hugging Face counts the requests for a file, not the people behind them, so one busy server can add a lot of downloads. Read the totals as a sign of interest from developers. They aren't users.

How many people use DeepSeek?

DeepSeek had 129 million monthly active users on its app in China in June 2026, according to QuestMobile, up from 127 million in March. That puts it third among the AI apps in China, behind ByteDance's Doubao at 382 million and Alibaba's Qwen at 167 million. DeepSeek itself doesn't publish a count of its users.

People who open DeepSeek tend to stay. QuestMobile found that 30.0% of DeepSeek's users in June 2026 fell in its 10 minutes and over bracket for time spent, against 27.5% for Doubao and 26.1% for Kimi, per Jiemian.

Similarweb's numbers for the website are estimates from a model, not counts that came from DeepSeek. By its estimate, China sent 50.47% of the desktop visits to deepseek.com in September 2026, Russia sent 9.04% and the US sent 5.8%. The same page puts the audience at 63.31% male, and it says that 25 to 34 year olds are the largest group.

DeepSeek sends almost no one onward to other websites. StatCounter, which counts clicks from chatbots to the sites they cite, gave it 0.02% of those referrals worldwide in September 2026, against 80.48% for ChatGPT. That's close to zero. Pew didn't ask about DeepSeek in its February 2026 survey of US adults, so no survey team has a US usage rate for it. Our AI chatbot market share page sets these figures next to every rival.

How much did it cost to train DeepSeek?

DeepSeek put the final training run of V3 at $5.576 million and the training that turned V3 into R1 at $294,000. The V3 figure is the cost of renting Nvidia H800 chips at $2 an hour. Neither number covers the chips it bought, its staff or any of the research that came before. So neither is a budget.

The V3 technical report says it plainly: "Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M." The same passage says the sum leaves out "the costs associated with prior research and ablation experiments". V3 has 671 billion parameters, 37 billion of them active per token, and DeepSeek trained it on 14.8 trillion tokens.

The R1 number arrived later. DeepSeek gave it in a peer reviewed paper in Nature on 17 September 2025: 512 H800 chips, 80 hours for the final stage, and $294,000 in all, Reuters reported. The paper also admitted for the first time that DeepSeek owns Nvidia A100 chips, used to prepare experiments on a smaller model. For scale, Reuters noted that Sam Altman said in 2023 that training OpenAI's foundation models cost "much more" than $100 million.

What does the DeepSeek API cost in October 2026?

DeepSeek's Flash model costs $0.30 per million input tokens and $1.20 per million output tokens at peak, and half that off peak, per DeepSeek's pricing page on 12 October 2026. V4 Pro costs $1.32 and $3.96 at peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.

Prices are per million tokens. A standard answer is 1,000 tokens in and 500 out at the cache miss rate, the same sum behind our Model Cost Index. Off peak, a Flash answer costs $0.00045. That's 73% less than R1 at its January 2025 launch. V4 Pro off peak costs almost exactly what R1 did then. Rivals' rates are in LLM API pricing statistics.

DeepSeek has also published its own margin. Over 24 hours on 27 and 28 February 2025, its V3 and R1 services handled 608 billion input tokens and 168 billion output tokens. They ran on an average of 226.75 server nodes, at a cost of $87,072, per its open infrastructure notes. Billed at R1 prices, that day would have brought in $562,027, a cost profit margin of 545%. DeepSeek said real revenue is much lower, because the app and website are free, V3 costs less than R1, and night traffic is discounted.

You don't have to buy from DeepSeek to use its models. On 12 October 2026, 18 other providers served V4 Pro on OpenRouter, Together, Cloudflare, Baidu and Alibaba among them, at $0.25 to $1.65 per million input tokens. Our DeepSeek server busy guide covers when that matters.

How much is DeepSeek worth?

DeepSeek is worth about $51.8 billion, the valuation implied by a Chinese company filing that Reuters reported on 16 July 2026. A fund paid 2.9 billion yuan for an indirect stake of 0.8265%, which values DeepSeek at 350.88 billion yuan. Reuters warned that this is one deal's price, not an announced round valuation.

It was the first time DeepSeek took money from outside investors. Reuters reported in June 2026 that DeepSeek was raising about 50 billion yuan ($7.4 billion) from backers that included Tencent and CATL. Reuters later reported plans for a further round at about 500 billion yuan and early work on a listing on Shanghai's STAR Market. DeepSeek hasn't confirmed any of it, and as a private company it doesn't have to.

Its biggest effect on the markets came long before it had a price of its own. On 27 January 2025, after the free DeepSeek app overtook ChatGPT in downloads on Apple's US App Store, Nvidia shares fell just under 17%. That erased $593 billion of market value, a one day record for any company on Wall Street, per Reuters and LSEG data.

Where is DeepSeek banned or restricted?

At least five governments or regulators acted against DeepSeek's app in the first half of 2025, and most of them did it over user data that was sent to China. Most of the orders cover government devices or app stores. It stayed legal for people in the US and Australia to use it on their own.

The data question comes from DeepSeek's own privacy policy, last updated on 10 February 2026: "we directly collect, process and store your Personal Data in People's Republic of China." In January 2025, Wiz Research found a DeepSeek database open to the internet with no login. It held over a million log lines, with chat history in plain text and API keys, and DeepSeek locked it after Wiz reported it. Is DeepSeek safe goes through the policy line by line.

If you want DeepSeek without a DeepSeek account, Whizi's Pro plan ($29.99 a month) has five DeepSeek models next to Claude and GPT. Using DeepSeek in Whizi lists them.

Where does each DeepSeek number come from?

We opened every figure at its source between 10 and 12 October 2026. Those are DeepSeek's papers, API docs and privacy policy, QuestMobile's reports as Chinese outlets carried them, Similarweb, StatCounter, Pew, the regulators, and Reuters where a source spoke to a reporter. Similarweb's figures are estimates, and we label them that way.

Two of the numbers are our own work. The 399 million is the sum of the all time download field for all 105 models in the deepseek-ai account on Hugging Face. We pulled it from the public API on 12 October 2026, and we grouped the models into the families in the chart. The cost of a standard answer is our own sum from DeepSeek's list prices.

We left out what we couldn't trace. That cut the "22.15 million daily active users" and "96.88 million monthly active users" that sit on many DeepSeek statistics pages. It also cut a "33.7 million monthly users" figure that is credited to Reuters, because we couldn't find it in any Reuters story. The one app download figure we kept is the install count on DeepSeek's own Google Play listing. Combined App Store and Google Play totals only appear on aggregators, so they're out.

How should you cite these DeepSeek statistics?

You're free to use any figure or table here with a link back to this page. Please credit the original source for numbers that came from someone else. A credit line like this one works:

Source: Whizi Research, "DeepSeek statistics: 399 million model downloads and 129 million users in China", whizi.io/resources/deepseek-statistics, updated October 2026.

More from the same series: Claude statistics, ChatGPT statistics and Gemini statistics.