Overview

DeepSeek is a Chinese artificial-intelligence company based in Hangzhou, Zhejiang, and backed and owned by the Chinese quantitative hedge fund High-Flyer. The company was established in July 2023 when an artificial-general-intelligence research group created by High-Flyer was separated into an independent business. Liang Wenfeng, High-Flyer's co-founder, founded DeepSeek and serves as its chief executive. The company was created to pursue general AI research outside High-Flyer's financial operations and has maintained a research-centered posture rather than presenting itself primarily as a conventional enterprise software vendor. DeepSeek develops large language models, coding models, mathematics models, mixture-of-experts systems, and consumer-facing conversational products. Its best-known releases include DeepSeek Coder, DeepSeek-LLM, DeepSeek-MoE, DeepSeek-Math, the V2 and V3 families, and DeepSeek-R1. The company has generally distributed model weights, code, or model-access services under licenses that vary by release. The term open-weight is more precise than fully open source: model parameters may be made available, while training data and every element of the training process are not necessarily openly licensed or reproducible. DeepSeek attracted international attention because its models combined competitive benchmark performance with comparatively economical training and inference requirements. The company has emphasized algorithmic efficiency, mixture-of-experts routing, reduced memory use, and the use of available computing hardware. Its research and engineering work developed alongside US restrictions on the export of advanced semiconductor products to China, encouraging optimization for constrained hardware resources. DeepSeek has also operated substantial internal computing infrastructure, including the Fire-Flyer and Fire-Flyer 2 clusters associated with High-Flyer. The company’s international profile expanded sharply after the January 2025 release of DeepSeek-R1 and an accompanying chatbot. R1 demonstrated strong reasoning performance and was compared with leading proprietary systems such as OpenAI’s GPT-4 and o1. Its availability through a consumer app, web services, APIs, and downloadable model variants made DeepSeek a prominent alternative to closed commercial AI platforms. The launch also intensified debate about the cost of frontier-model development, the strategic implications of open-weight AI, US-China technology competition, data governance, censorship, and the relationship between commercial AI research and national-security institutions. DeepSeek’s positioning combines high model capability, comparatively low cost, open or accessible distribution, and research-led development. It has reportedly recruited heavily from Chinese universities and has hired researchers from fields beyond traditional computer science, including mathematics and the humanities. The company has stated that it focuses on research and has not prioritized immediate commercialization, although its chatbot, API access, model hosting, and developer ecosystem give it a practical presence in the global AI market. Its current status is active and private.

History

DeepSeek’s origins lie in High-Flyer, a Chinese quantitative investment firm co-founded by Liang Wenfeng. Liang became interested in machine learning during his studies at Zhejiang University and used increasingly sophisticated computational methods in High-Flyer’s trading activities. The firm began using a GPU-dependent deep-learning trading system in 2016, and by the end of the following year most of its trading activity was reportedly driven by artificial-intelligence systems. High-Flyer subsequently developed its own computing capacity. In 2019 it began building the Fire-Flyer cluster, which used more than one thousand GPUs connected through a high-speed network. That system was retired after roughly a year and a half. A larger successor, Fire-Flyer 2, began construction in 2021. Its infrastructure included thousands of Nvidia A100 GPUs and software designed for scheduling, distributed training, communication, storage, fault handling, and recovery. The computing environment became an important foundation for High-Flyer’s later AI research. In April 2023, High-Flyer announced an artificial-general-intelligence research laboratory intended to work on AI tools separate from the hedge fund’s financial business. The laboratory was spun out in July 2023 as DeepSeek, with High-Flyer serving as its principal financial backer. Venture investors were reportedly cautious because a research-oriented AI company did not offer an obvious short-term exit. DeepSeek nevertheless continued to build a large research and engineering team in Hangzhou under Liang’s leadership. The company’s early model releases concentrated on programming and general language tasks. DeepSeek Coder appeared in November 2023, followed later that month by DeepSeek-LLM. In 2024, DeepSeek expanded into mixture-of-experts models and mathematical reasoning systems, then released DeepSeek-V2 and DeepSeek-V2.5. These models introduced or refined architectural techniques such as multi-head latent attention and sparse expert routing, which can reduce memory or computation requirements compared with conventional dense transformer systems. DeepSeek-V3 and its later revisions continued this efficiency-oriented approach. The company reported a relatively low training cost for V3 compared with figures publicly associated with some competing frontier models, although such comparisons depend on accounting definitions, hardware assumptions, engineering labor, data preparation, and whether research and infrastructure costs are included. The company’s work also took place under US restrictions affecting China’s access to certain advanced chips. This environment encouraged the use of software optimization, older or less capable export-compliant hardware, and training methods intended to extract more performance from limited compute. The November 2024 preview of DeepSeek-R1-Lite and the January 2025 release of DeepSeek-R1 marked the company’s most important public expansion. R1 was a reasoning-oriented model whose answers were reported to be competitive with contemporary proprietary systems. DeepSeek released the model through online services and applications and also provided downloadable variants, including smaller distilled models. The launch made DeepSeek one of the most visible Chinese AI brands internationally and prompted renewed debate about whether frontier AI development necessarily required the spending and hardware scale associated with leading US companies. DeepSeek’s model distribution strategy has combined public weights, permissive licenses for some later releases, hosted chat access, and APIs. This has encouraged experimentation by developers but has also generated debate over whether open-weight systems should be described as open source when their training data and full development pipeline are not public. The company has also faced scrutiny concerning censorship, privacy, security, Chinese regulatory obligations, and reported links between some researchers and Chinese military or defense-related institutions. These issues became part of the broader geopolitical discussion surrounding DeepSeek’s rise. The company has continued to emphasize research, technical efficiency, and recruitment based on demonstrated ability rather than long professional experience. Its hiring has included graduates and people from nontraditional technical backgrounds, including mathematics and language-related disciplines. DeepSeek has also drawn attention in emerging markets, including Africa, where lower-cost and less energy-intensive systems and the possibility of local data control are attractive. It remains a private company backed by High-Flyer and is best understood as an AI research and model-development organization whose chatbot and model releases have become its principal public touchpoints.

2025
DeepSeek-R1 and the DeepSeek chatbot launch
[1]DeepSeek - Wikipediaen.wikipedia.org
DeepSeek released R1 and an associated chatbot in January, bringing the company to global public attention.
2025
DeepSeek releases V3-0324 under the MIT License
[1]DeepSeek - Wikipediaen.wikipedia.org
The company published a revised V3 model under the MIT License.
2025
DeepSeek-R1-0528 is released
[1]DeepSeek - Wikipediaen.wikipedia.org
DeepSeek issued an updated R1 model with reported gains on selected evaluations.
2024
DeepSeek expands into MoE and mathematics models
[1]DeepSeek - Wikipediaen.wikipedia.org
DeepSeek released mixture-of-experts models and a family of mathematics-focused systems.
2024
DeepSeek-V2 and V2.5 are introduced
[1]DeepSeek - Wikipediaen.wikipedia.org
The V2 generation advanced DeepSeek’s use of efficient attention and expert-routing techniques, followed by the V2.5 revision.
2024
DeepSeek-R1-Lite preview becomes available
[1]DeepSeek - Wikipediaen.wikipedia.org
A preview of the company’s reasoning-oriented R1-Lite system became available through online access.
2023
DeepSeek is established as an independent company
[1]DeepSeek - Wikipediaen.wikipedia.org
High-Flyer’s AGI research laboratory was spun out as DeepSeek in July 2023, with Liang Wenfeng as founder and chief executive.
2023
DeepSeek Coder is released
[1]DeepSeek - Wikipediaen.wikipedia.org
DeepSeek introduced its first publicly documented model family, focused on code generation and understanding.
2023
DeepSeek-LLM is released
[1]DeepSeek - Wikipediaen.wikipedia.org
The company expanded from coding systems into general-purpose language models in base and chat forms.
2021
Fire-Flyer 2 project begins
[1]DeepSeek - Wikipediaen.wikipedia.org
Construction began on the larger Fire-Flyer 2 cluster, which later supported large-scale model training.
2019
Fire-Flyer computing cluster begins construction
[1]DeepSeek - Wikipediaen.wikipedia.org
High-Flyer began constructing its first major AI computing cluster, later known as Fire-Flyer.
2016
High-Flyer begins GPU-based deep-learning trading
[1]DeepSeek - Wikipediaen.wikipedia.org
High-Flyer began using a GPU-dependent deep-learning model in its trading operations.
2015
High-Flyer is co-founded
[1]DeepSeek - Wikipediaen.wikipedia.org
Liang Wenfeng co-founded High-Flyer, the quantitative investment firm that later financed and incubated DeepSeek.

Products and positioning

Research-led provider of capable, computationally efficient, and comparatively affordable open-weight AI models.

Code-generation model · 2023
DeepSeek Coder
DeepSeek Coder is a family of code-oriented language models released in base and instruction-tuned forms. It was trained primarily on source code and related technical text and supports programming assistance, code completion, explanation, and transformation. The family established DeepSeek’s initial public identity as a model developer and was later extended through the Coder V2 series.
General-purpose language model · 2023
DeepSeek-LLM
DeepSeek-LLM is a general language-model series offered in base and chat variants. Its initial releases included 7-billion- and 67-billion-parameter models trained on English and Chinese text. The series broadened DeepSeek’s scope from programming systems to general text generation, question answering, and conversational use.
Mixture-of-experts model · 2024
DeepSeek-MoE
DeepSeek-MoE models use sparse mixture-of-experts routing so that only selected expert components process each token. This approach is intended to increase model capacity without requiring every parameter to be active for every input, supporting DeepSeek’s broader emphasis on computational efficiency.
Mathematics reasoning model · 2024
DeepSeek-Math
DeepSeek-Math is a family of models designed for mathematical problem solving and reasoning. The series extended the company’s research beyond general chat and coding, and was used to demonstrate capabilities on mathematical benchmarks and structured reasoning tasks.
Large language model · 2024
DeepSeek-V2
DeepSeek-V2 introduced a new generation of general language models using techniques including multi-head latent attention and mixture-of-experts architecture. The design sought to reduce key-value cache memory use and improve the economics of serving long-context or high-volume workloads.
Large language model · 2024
DeepSeek-V3
DeepSeek-V3 is a later general-purpose model family built around sparse expert routing and other efficiency-focused engineering choices. The model became a central part of DeepSeek’s public technology story because the company reported strong performance while claiming substantially lower training expenditure than some competing frontier systems.
Reasoning language model · 2025
DeepSeek-R1
DeepSeek-R1 is a reasoning-oriented model released in January 2025. It became DeepSeek’s signature product after demonstrating competitive results on several reasoning and general-language evaluations. DeepSeek also released smaller distilled variants, making the R1 approach available across a range of hardware and deployment requirements.
AI conversational assistant · 2025
DeepSeek chatbot
The DeepSeek chatbot provides a consumer-facing interface to the company’s language and reasoning models through web and mobile experiences. Its launch alongside R1 made DeepSeek directly accessible to a global audience and helped turn the company’s model research into a recognizable consumer brand.

Controversies

2026
Anthropic allegation concerning Claude-generated training data
Anthropic accused DeepSeek of using thousands of fraudulent accounts to generate large volumes of conversations with Claude for possible model training. The claim is reported as an allegation and should not be treated as a proven violation without further authoritative findings.
Source
2026
Anthropic report alleges Alibaba, Moonshot AI and DeepSeek used Claude outputs without authorization to train their own models
Anthropic said in a threat intelligence report released on September 10, 2026 that it detected and disrupted unauthorized large-scale "distillation" of its Claude models by China-based AI labs including Alibaba, Moonshot AI and DeepSeek, which allegedly used Claude outputs to train and improve their own models such as Qwen and Kimi.
Source
2026
U.S. restricts DeepSeek technology in federal civilian agencies
In 2026, the U.S. Department of Homeland Security issued an order requiring federal civilian agencies to remove DeepSeek-related technology from their IT systems and barring them from procuring DeepSeek products and services. According to the order, the stated concerns were potential access to sensitive government data and cybersecurity vulnerabilities. The directive applies to federal civilian agencies and covers both the removal of existing DeepSeek-related technology from agency IT systems and a prohibition on purchasing DeepSeek products and services. The order did not, as described in the available summary, specify further measures beyond these requirements.
Source
2025
Security, privacy, censorship, and defense-affiliation scrutiny
DeepSeek’s rapid international growth prompted scrutiny over data handling, privacy, content controls, Chinese regulatory obligations, and reported relationships between some researchers and Chinese military or defense-related institutions. These concerns formed part of wider geopolitical debate about Chinese AI systems, but the available reference material does not establish a single adjudicated scandal or final legal determination.
Source

Brand decisions

2025
Release a high-performing reasoning model with broad accessProduct launch
Background
DeepSeek had developed a reasoning-oriented model and sought to make its research available beyond private internal use.
Action taken
The company released DeepSeek-R1 through applications, web access, APIs, and downloadable variants.
Aftermath
R1 rapidly increased DeepSeek’s international visibility and intensified debate about open-weight AI and the economics of frontier-model development.
Source
2025
Publish revised V3 and R1 models under permissive licensingGeneration change
Background
After earlier releases under the company’s own or source-available licenses, DeepSeek sought wider downstream experimentation.
Action taken
The company released V3-0324 and R1-0528 under the MIT License.
Aftermath
The releases expanded the availability of DeepSeek technology while continuing debate over the distinction between open weights and fully open-source AI.
Source
2024
Prioritize research and computational efficiencyStrategy
Background
Access to advanced computing hardware was constrained by US semiconductor export controls, while frontier-model training required substantial resources.
Action taken
DeepSeek emphasized algorithmic optimization, sparse expert architectures, memory reduction, and efficient use of available GPU infrastructure.
Aftermath
The company became associated with high reported benchmark performance at comparatively low claimed training and inference costs.
Source
2023
Separate AGI research from High-Flyer’s financial businessStrategy
Background
High-Flyer had developed substantial AI research and computing capabilities but wanted a laboratory focused on general AI rather than financial trading.
Action taken
The research laboratory announced in April 2023 was spun out as the independent DeepSeek company in July.
Aftermath
DeepSeek retained High-Flyer as its principal backer while developing models under its own brand.
Source

Recent events

2026.10.31
DeepSeek publishes Super-agent architecture[6]DeepSeek发布Super-agent架构hkcd.com
DeepSeek published a technical article describing its Super-agent form and system architecture. The article says an internal instance built on these ideas is already running continuously.
2026.10.31
DeepSeek publishes Super-agent architecture[7]DeepSeek发布Super-agent架构yicai.com
DeepSeek's Harness team published a technical article outlining its Super-agent architecture and design. The article says an internal instance built on the approach is already running.
Unconfirmed
Reported2026.10.23
DeepSeek and MiniMax receive approval for global AI use[8]DeepSeek与MiniMax获准在全球使用AInews.google.com
Reuters reported that DeepSeek and MiniMax won approval for AI global use, citing a state media-affiliated entity.
Unconfirmed
Reported2026.10.22
DeepSeek raises approximately RMB 100 billion[9]DeepSeek融资约1000亿元zaobao.com.sg
Bloomberg reported that DeepSeek raised approximately RMB 100 billion in a new funding round, with CATL and Tencent among the largest investors.
Unconfirmed
DeepSeek open-sources DeepInfer[10]DeepSeek 开源 DeepInferglobaltimes.cn
DeepSeek announced that it had open-sourced DeepInfer, its internal inference framework, to help developers and enterprises optimize inference throughput and latency and lower compute costs and energy consumption.
Unconfirmed
DeepSeek announces usage-based fees for web search and fetch API calls[11]DeepSeek宣布网页搜索与抓取API调用按次收费chinaz.com
DeepSeek announced that API calls to its web search and fetch interfaces would be charged at 5 yuan per 1,000 successful requests after a free transition period, starting November 1 at 00:00.
2026.10.19
Reported2026.10.20
U.S. DHS orders DeepSeek products and services excluded from federal civilian procurement[12]美国国土安全部下令将DeepSeek产品与服务排除出联邦民用采购finance.sina.com.cn
The U.S. Department of Homeland Security issued an exclusion and removal order covering DeepSeek and affiliated entities’ products and services in federal civilian agencies’ procurement channels and IT systems.
2026.10.17
Reported2026.10.18
DeepSeek updates V4.1 Flash[13]DeepSeek更新V4.1 Flashcqnews.net
DeepSeek said it updated V4.1 Flash to V4.1-Flash-1017 and synchronized the new model across its official website, app, and API. DeepSeek Harness 0.2.1 was adapted to the new model.
2026.10.17
Reported2026.10.17
DeepSeek expands access to Search and Fetch APIs[14]DeepSeek扩大Search和Fetch API的使用范围econotimes.com
DeepSeek expanded access to its Search and Fetch APIs, making both services available to all users. The APIs are priced at $0.75 per 1,000 calls for each endpoint.
2026.10.17
Reported2026.10.17
DeepSeek announces V4.1 Flash model update[14]DeepSeek扩大Search和Fetch API的使用范围econotimes.com
DeepSeek announced the release of DeepSeek-V4.1-Flash-1017, an updated AI model with improved performance and cost efficiency, available through its website, mobile application, and API.