DeepSeek is often described with numbers that measure different things: downloads, active users, benchmark results, training compute, and revenue. Each figure below includes its metric and period. API prices were checked against the official pricing page on October 7, 2026.
- QuestMobile put DeepSeek's active-user scale in China's native AI apps at 129 million in June 2026. Sensor Tower reported 127 million worldwide mobile app downloads by the end of June 2025. These are different measures.
- DeepSeek's V3 report lists 2.788 million H800 GPU hours. At an assumed $2 per GPU hour, that works out to $5.576 million. Earlier research and ablation experiments are excluded.
- DeepSeek released V4.1 Flash on September 10, 2026. Its benchmarks are provider results. Reuters reported on October 6 that another funding round had not yet closed.
1. Company and origins
DeepSeek was founded in Hangzhou in 2023. Its CEO, Liang Wenfeng, had previously co-founded the quantitative hedge fund High-Flyer. These details come from an Associated Press report.
2. Downloads and user numbers
The available figures cannot be combined into a single user total. Sensor Tower tracks mobile app downloads. QuestMobile reports an active-user scale within its market for native AI apps in China.
The first two figures are from Sensor Tower's report on the app's launch and its later first-half 2025 review. In its half-year report dated July 14, 2026, QuestMobile gives DeepSeek an active-user scale of 129 million in June. The figure concerns China's native AI app market. The source calls it an active-user scale, not worldwide accounts or downloads.
Sensor Tower also reported that DeepSeek reached number one in the US App Store on January 27, 2025, and in Google Play in the US the following day. For the first 19 days, the report gives the download shares shown in the table below. They refer to downloads, not monthly active users.
Sensor Tower published these geographic shares in its February 2025 report. It does not provide a complete country breakdown. The table therefore does not calculate an "Other" category.
3. Models and published benchmarks
Benchmark results depend on model version, prompt, compute budget, and test method. The figures below are historical numbers from DeepSeek's own reports. They show what the provider published, not an independent ranking.
3.1 DeepSeek-V3: training and original comparison
The DeepSeek-V3 technical report lists 671 billion parameters in total and 37 billion activated per token. Pretraining used 14.8 trillion tokens. The report records 2.788 million H800 GPU hours across pretraining, context extension, and post-training. At an assumed rental price of $2 per GPU hour, that works out to $5.576 million. The report excludes earlier research and ablation experiments. It is therefore not the company's full development spending.
The table reproduces selected results from the original DeepSeek-V3 study. It compares the provider's figures with the older GPT-4o and Claude 3.5 Sonnet results listed in that report. The scores apply to the named tasks and metrics.
For the DeepSeek-V3 base model, the same report separately lists 87.5% on BBH (exact match, 3-shot). That figure comes from the base-model evaluation and is not part of the chat-model comparison in the table.
3.2 Reasoning models
DeepSeek’s R1 model card reports 90.8% on MMLU (Pass@1) and 79.8% on AIME 2024 (Pass@1). The MMLU figure uses a different metric from the V3 MMLU exact-match score above. The R1-0528 release note compares the earlier R1 at 70.0% with R1-0528 at 87.5% on AIME 2025. The 87.5% figure belongs to R1-0528, not the original R1.
The R1-0528 result appears in the May 28, 2025 release note. The V3.2 figures come from DeepSeek's technical report. On the AIME 2025 row, the report pairs V3.2-Speciale's score with 23,000 generated tokens and GPT-5 High's with 13,000 generated tokens. These should not be read as scores at the same output budget. All entries in this table are provider results.
3.3 DeepSeek-V4 and V4.1 Flash
DeepSeek previewed V4 on April 24, 2026. The official V4 report describes V4 Pro with 1.6 trillion parameters, 49 billion of them active. It lists V4 Flash at 284 billion parameters, with 13 billion active. Both versions support a context window of up to 1 million tokens, according to the report. DeepSeek reports the V4 Pro benchmark results below for max mode. These are provider results, not an independent rerun.
In max mode, DeepSeek reported 87.5% on MMLU-Pro and 90.1% on GPQA Diamond for V4 Pro. The provider gave a 93.5% LiveCodeBench result. These figures also come from the V4 report.
V4 Pro became generally available on August 13, 2026. DeepSeek released V4.1 Flash on September 10. Its model card describes a multimodal Mixture-of-Experts model with image processing and a context window of up to 1 million tokens. It activates 8 billion at prefill and 16 billion at decode. According to the September 10 release note, V4 Flash and V4 Vision Experimental were retired. Legacy API calls under the V4 Flash name are routed to V4.1 Flash.
These figures come from the official V4.1 Flash model card. DeepSeek ran most tasks at maximum reasoning effort. Agent benchmarks used DeepSeek Harness Minimal and a one-million-token context limit; named test frameworks were used for some tasks. The results are provider claims. The model card lists 39.1% for the text-only subset of HLE, identified in its table footnote.
4. API pricing
The pricing page lists V4.1 Flash under API model ID “deepseek-flash” and V4 Pro as “deepseek-v4-pro”. The table reproduces the rates listed there on October 7, 2026, in US dollars per one million tokens. It distinguishes cache hits, uncached input, output, and peak versus off-peak periods. Values come from the project’s centralized model data and were checked against the official pricing page.
DeepSeek defines one weekday peak window from 01:00 to 04:00 UTC. A second runs from 06:00 to 10:00 UTC. Weekends and Chinese public holidays use off-peak rates, according to the pricing page.
The September 10 launch note initially said that requests using “deepseek-v4-pro” would route to V4.1 Flash at Flash rates from September 14 until V4.1 Pro launches. A later entry in the official changelog says DeepSeek would continue Pro API service after September 14 at users’ request, with billing unchanged. The current pricing page also lists V4 Pro with its own model version and rates. The provider’s documents therefore confirm continued availability and list those prices, but do not explain in detail how the initially announced routing change was implemented. Individual usage invoices are not public.
5. Revenue and funding
The sources reviewed do not include audited annual accounts published by DeepSeek. The following business figures come from media reports and should be distinguished from confirmed company disclosures.
IT Home reported the January through July figures on August 26, 2026, citing The Information and two people familiar with the financial data. They are not audited accounts. On September 24, Reuters reported that The Information put DeepSeek’s annualized revenue run rate at $1 billion. The figure came from two sources and an investor meeting; Reuters could not independently verify it. A run rate extrapolates the current pace across a year. It is not revenue earned over a full year and is not directly comparable with the 475 million yuan reported for January through July. The Information attributed some growth to price increases of 2.3 to 4.5 times. On October 6, 2026, Reuters reported, citing one person with knowledge of the round, that commitments exceeded 80 billion yuan, above an original 50 billion yuan target. Reuters gave a target valuation of 500 billion yuan, or about $74 billion, and said a close was expected in October. The report did not say the round had closed.
Reuters reported on June 3 that two unnamed sources estimated a planned round of about 50 billion yuan at a post-money valuation of 350 billion to 400 billion yuan. On July 15, other Reuters sources cited about 450 billion yuan. The next day, Reuters derived a transaction-implied company value of 350.88 billion yuan from a filing by Anhui Korrun. Reuters said the filing did not specify whether the investment involved new or existing shares or different rights. The figure therefore is not a definitive valuation for the funding round. Caixin Global also reported more than 350 billion yuan based on the filing. The sources do not establish one agreed round valuation. The amount raised is a separate measure.
6. Market reaction in January 2025
On January 27, 2025, NVIDIA shares fell 17% in one trading day, according to the Associated Press. Reuters reported at the time that the market-capitalization loss was on track to reach about $600 billion. Its report connected the move with investor concerns about DeepSeek’s reported model efficiency. This is a contemporaneous market interpretation, not proof that one announcement alone caused the entire share-price move.
7. Availability, model weights, and privacy
The DeepSeek app is listed in the German Apple App Store and on Google Play. DeepSeek also publishes a web chat and an API, as well as model weights on Hugging Face. Store pages confirm a listing, not successful registration or unchanged availability of every feature.
DeepSeek says in its privacy policy that it directly collects, processes, and stores personal information in the People’s Republic of China. This is the provider’s statement, not a legal assessment.
The official model repositories for V3, R1, V4 Pro, and V4.1 Flash specify the MIT license. “Open model weights” describes these releases more precisely than saying the entire DeepSeek service is open source. Whether a model can run locally also depends on hardware, software, and model version.
When weights run on your own device without an external inference service, prompts do not need to be sent to DeepSeek. Data flows still depend on the local runtime, logging, and connected services.
8. DeepSeek timeline
The official DeepSeek API news page lists the dated V3, app, R1, V3.2, and V4.1 releases. The January 15, 2025 app announcement, V3.2-Exp release note, V4-Flash-0731 changelog, R1-0528 note, V4 preview announcement, and V4 Pro GA announcement document other milestones.
For more statistics, see our articles on Claude, ChatGPT, and Gemini.
