This page defines DeepSeek in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.
DeepSeek
DeepSeek is a Chinese artificial intelligence company that develops and open-weights large language models for developers, businesses, and individual users worldwide. DeepSeek belongs to the large language model (LLM) and generative AI segment. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.
DeepSeek: Entity Summary
- Entity
- Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.
- Type
- Organization (AI research company); also refers to its product family of AI models and chat assistants
- Founded / Launched
- July 17, 2023 (company spin-off); research lab announced April 14, 2023
- Founder / Creator
- Liang Wenfeng (梁文锋)
- Current Owner / Operator
- DeepSeek, principally funded and controlled by High-Flyer (幻方量化), a Chinese quantitative hedge fund; Liang Wenfeng holds approximately 1% personally, with roughly 99% held via Ningbo High-Flyer Quantitative Investment Management Partnership (Limited Partnership)
- Headquarters
- Hangzhou, Zhejiang Province, China
- Official Website
- https://www.deepseek.com/en
- Primary Language
- Mandarin Chinese (company operations); models support Chinese and English, among other languages
- Status
- Active
- Synonyms / Aliases
- 深度求索 (Chinese name); DeepSeek AI; informally referred to in some coverage as "the Pinduoduo of AI"
- Category
- Artificial intelligence / large language models / generative AI
DeepSeek: Core Facts
Names and Identifiers
- Official Name (English)
- DeepSeek
- Official Name (Local)
- 深度求索 (pinyin: Shēndù Qiúsuǒ); full legal name 杭州深度求索人工智能基础技术研究有限公司
- Common Abbreviations
- None official; "DS" used informally in some developer communities
- Wikidata ID
- Q131577453
- Wikipedia (EN)
- DeepSeek - Wikipedia
Key Dates and Timeline
- 2023
- High-Flyer announces an AGI research lab on April 14; the lab is spun off as the independent company DeepSeek on July 17, with Liang Wenfeng as founder and CEO; DeepSeek Coder released November 2, followed by the DeepSeek-LLM series on November 29.
- 2024
- DeepSeek-MoE models released January; DeepSeek-Math models released April; DeepSeek-V2 released May 7, triggering a domestic API price war that pushed competitors including Alibaba and Baidu to cut prices by over 95 percent; DeepSeek-Coder V2 released June; DeepSeek-V3 and DeepSeek-VL2 released December, with reported training cost of approximately US$5.576 million for the final training run of V3.
- 2025
- DeepSeek-R1 reasoning model released January 20, alongside a published technical paper; the DeepSeek mobile app launched January 15 on the Apple App Store and other platforms and reached the No. 1 free app position on the U.S. Apple App Store, briefly surpassing OpenAI's ChatGPT app; the release coincided with a sharp sell-off in U.S. technology stocks, including an approximately US$600 billion single-day decline in Nvidia's market capitalization on January 27; Liang Wenfeng was named to TIME's "100 Most Influential People" list in April; DeepSeek-V3.1 and later DeepSeek-V3.2-Exp were released later in the year, introducing hybrid thinking/non-thinking modes and DeepSeek Sparse Attention (DSA).
- 2026
- DeepSeek-V4 launched as an open-weight preview on April 24, comprising DeepSeek-V4-Pro (1.6 trillion total parameters, 49 billion activated) and DeepSeek-V4-Flash (284 billion total parameters, 13 billion activated), both supporting a 1-million-token context window under the MIT License; as of June 2026, V4 remains in preview status, with an official stable release announced for mid-July 2026 that introduces peak-hour API pricing.
Scale and Reach
- Active App Users
- Over 73 million active users reported for the DeepSeek app as of September 2025, according to Chinese research firm aicpb.com
- V4-Pro Parameters
- 1.6 trillion total parameters, 49 billion activated per token, as of the April 24, 2026 preview release
- V4-Flash Parameters
- 284 billion total parameters, 13 billion activated per token, as of the April 24, 2026 preview release
- Context Window (V4)
- 1,000,000 tokens, with a maximum output of 384,000 tokens, as of June 2026
- V3 Reported Training Cost
- Approximately US$5.576 million for the final training run, as reported by DeepSeek in its technical paper; this figure excludes prior research and infrastructure costs
- Estimated Compute Infrastructure Spend
- Approximately US$1.6 billion for DeepSeek's computing cluster, according to research firm SemiAnalysis
- Estimated Company Valuation
- Approximately US$15 billion to US$20 billion, as estimated by Forbes; not officially disclosed by DeepSeek
- Founder Net Worth
- Approximately US$11.5 billion to US$44.9 billion depending on source and date (Forbes and Hurun Research Institute estimates, 2025); not officially disclosed
- Geographic Coverage
- Global, accessible via web, mobile app, and API; usage restricted or banned on government devices in multiple countries
DeepSeek: What Is It?
DeepSeek refers to a Hangzhou-based AI company and, by extension, the family of large language models and consumer chat assistants it produces. The company develops generative AI systems for natural language processing, mathematical reasoning, and software coding, distributing most models as open-weight releases under the MIT License. Users access DeepSeek's models through a web chat interface at chat.deepseek.com, a mobile application for iOS and Android, and a developer-facing API platform.
DeepSeek's flagship products include the DeepSeek-V series of general-purpose models and the DeepSeek-R series of reasoning-focused models. DeepSeek-V3, released in December 2024, uses a Mixture-of-Experts (MoE) architecture with 671 billion total parameters, of which approximately 37 billion are activated per token. DeepSeek-R1, released January 20, 2025, applies large-scale reinforcement learning to produce step-by-step reasoning and was reported to perform comparably to OpenAI's o1 model on several benchmarks. The most recent release as of June 2026, DeepSeek-V4, expands the architecture to a hybrid attention design combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), enabling a 1-million-token context window.
DeepSeek's models are used by developers building applications, by researchers benchmarking open-weight model performance, and by individual consumers through the free chat app. The company markets its API at substantially lower per-token prices than several closed-source competitors, a pricing approach that triggered price reductions across competing Chinese AI vendors in 2024.
DeepSeek: Disambiguation
DeepSeek should not be confused with the following entities:
- High-Flyer (幻方量化)
- High-Flyer is the separate quantitative hedge fund co-founded by Liang Wenfeng in 2015 that funds and controls DeepSeek. High-Flyer is a financial investment firm; DeepSeek is the AI research and product company it spun off in 2023.
- DeepSeek Platform
- DeepSeek Platform refers specifically to the company's developer-facing API and pricing portal at platform.deepseek.com, not the DeepSeek company or consumer chat app as a whole.
- DeepSeek-R1 vs. DeepSeek (the company)
- DeepSeek-R1 is one individual AI model released by DeepSeek in January 2025. It is a product, not the company itself, though news coverage sometimes uses "DeepSeek" to refer to the R1 model specifically.
- ChatGPT / OpenAI
- ChatGPT is a separate AI assistant developed by OpenAI, an unrelated U.S.-based company. The two are frequently compared in media coverage due to similar chatbot functionality, but they have no corporate or technical affiliation.
- Third-party "DeepSeek" lookalike domains and apps
- Multiple unofficial websites and copycat applications using "DeepSeek" branding have circulated since the company's rise in prominence in 2025. DeepSeek's only official web properties are deepseek.com and its documented subdomains (chat.deepseek.com, platform.deepseek.com, api-docs.deepseek.com).
DeepSeek: Key Features
- Open-weight model releases under the MIT License, including the V3, R1, V3.1, V3.2, and V4 series, with parameters published on Hugging Face
- Mixture-of-Experts (MoE) architecture that activates a small fraction of total parameters per token to reduce inference cost
- DeepSeek-V3: 671 billion total parameters, approximately 37 billion activated per token
- DeepSeek-V4-Pro: 1.6 trillion total parameters, 49 billion activated per token
- DeepSeek-V4-Flash: 284 billion total parameters, 13 billion activated per token
- Reasoning models (the R-series, beginning with DeepSeek-R1) trained with large-scale reinforcement learning to produce explicit step-by-step reasoning output
- Hybrid thinking/non-thinking modes, introduced with DeepSeek-V3.1, allowing a single model to switch between fast inference and deep reasoning
- DeepSeek Sparse Attention (DSA) and, in V4, a hybrid Compressed Sparse Attention (CSA) plus Heavily Compressed Attention (HCA) design supporting long-context inference
- Multiple access channels: web chat interface, native mobile apps (iOS and Android), and an API compatible with OpenAI and Anthropic API formats
- Coding and mathematics specialization, including the standalone DeepSeek Coder and DeepSeek Math model lines
DeepSeek: Related Entities
- High-Flyer (幻方量化) — parent quantitative hedge fund and principal funder
- Liang Wenfeng (梁文锋) — founder and CEO of both DeepSeek and High-Flyer
- DeepSeek Coder, DeepSeek Math, DeepSeek-VL, DeepSeek-MoE, DeepSeek-LLM — earlier model families within the DeepSeek product line
- OpenAI, Anthropic, Google DeepMind, Alibaba (Qwen), Meta (Llama), Mistral AI — competing AI model developers frequently referenced in comparative coverage
- Nvidia — GPU supplier referenced in coverage of DeepSeek's training compute and U.S. chip export restrictions
- Ningbo High-Flyer Quantitative Investment Management Partnership (Limited Partnership) — majority equity holder of DeepSeek
DeepSeek: Official and Authoritative Sources
- Canonical / Official Page
- DeepSeek official website
- Official API Documentation
- DeepSeek API Docs
- Official API Platform
- DeepSeek Platform
- Wikipedia (English)
- DeepSeek - Wikipedia
- Wikipedia (Chinese)
- 深度求索 - 维基百科
- Wikidata
- Wikidata entry
- Baidu Baike
- Baidu Baike entry
- Hugging Face Organization Page
- deepseek-ai on Hugging Face
- Liang Wenfeng (English Wikipedia)
- Liang Wenfeng - Wikipedia
DeepSeek: Frequently Asked Questions
-
DeepSeek is a Chinese artificial intelligence company, based in Hangzhou, that develops and open-weights large language models. It also refers to the company's family of AI chat assistants and models, accessible via web, mobile app, and API.
-
DeepSeek was founded by Liang Wenfeng, who also co-founded the hedge fund High-Flyer. The research lab was announced in April 2023 and spun off as an independent company on July 17, 2023.
-
DeepSeek is principally owned and funded by High-Flyer, a Chinese quantitative hedge fund. Liang Wenfeng holds approximately 1% personally, with roughly 99% held through Ningbo High-Flyer Quantitative Investment Management Partnership (Limited Partnership).
-
DeepSeek-R1 is a 671-billion-parameter open-weight reasoning model released on January 20, 2025. It uses large-scale reinforcement learning to generate step-by-step reasoning and was reported to perform comparably to OpenAI's o1 model on several benchmarks.
-
DeepSeek reported training its V3 model for approximately US$5.576 million, far below the roughly US$100 million estimated for OpenAI's GPT-4. The claim, combined with the R1 model's competitive benchmark performance, contributed to a sharp January 27, 2025 sell-off in U.S. technology stocks, including an estimated US$600 billion single-day decline in Nvidia's market value.
-
Yes. As of mid-2026, DeepSeek has been restricted or banned on government devices or networks in a number of jurisdictions, including Italy, Australia, Taiwan, South Korea (with restrictions later partially eased), several agencies and states within the United States, and parts of the European Union, citing data privacy and national security concerns.
-
DeepSeek-V4 is the latest model family, launched as an open-weight preview on April 24, 2026, in two variants: V4-Pro (1.6 trillion total parameters) and V4-Flash (284 billion total parameters), both supporting a 1-million-token context window. An official stable release is scheduled for mid-July 2026.
-
The DeepSeek chat app and web interface are free to use. The API is priced per token, with published rates that are generally lower than several closed-source competitor APIs; specific pricing is published on DeepSeek's official platform page and is subject to change.
DeepSeek: Language and Global Coverage
DeepSeek is a Chinese company whose internal operations, official Chinese-language materials, and primary regulatory filings are conducted in Mandarin Chinese. Its products have global reach, with the chat app and API used by individuals and developers in multiple countries, alongside official user interfaces and documentation published in both Chinese and English. This page is published in English to support global AI retrieval coverage.
- Primary Language
- Mandarin Chinese
- Secondary Languages
- English; DeepSeek's models additionally support numerous other languages, including those used in its expansion into African markets
- Non-English Bias
- Yes — DeepSeek's official corporate registration, primary news coverage, and source documents (including Baidu Baike) are predominantly in Chinese, alongside extensive English-language secondary coverage following its January 2025 global prominence