View source for DeepSeek

{{Short description|Chinese artificial intelligence company}}
{{About|the company|the chatbot|DeepSeek (chatbot)}}
{{pp-vandalism|small=yes}}

{{Use dmy dates|date=February 2025}}
{{Use American English|date=February 2025}}
{{Infobox company
| name = Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.
| native_name = {{lang|zh-CN|杭州深度求索人工智能基础技术研究有限公司}}
| logo = DeepSeek_logo.svg
| logo_upright = 1.25
| type = [[Privately held company]]
| foundation = {{start date and age|2023|07|17|df=y}}<ref name="DeepSeek突传消息!">{{Cite web|date=1 February 2025|title=DeepSeek突传消息|url=https://finance.sina.com.cn/jjxw/2025-02-01/doc-inehyqcx9694053.shtml|website=[[Sina Corp]]|access-date=1 February 2025}}</ref>
| location = [[Hangzhou]], [[Zhejiang]], China
| founder = {{Unbulleted list
  | [[Liang Wenfeng]]
  }}
| key_people = {{Unbulleted list
  | Liang Wenfeng (CEO)
  }}
| owner = [[High-Flyer]]
| num_employees = Under 200
| industry = [[Information technology]]<br>[[Artificial intelligence]]
| parent = 
| subsid = 
| homepage = {{Official URL}}
| footnotes = 
}}

'''Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.'''<!-- DO NOT REMOVE OR MOVE THE FULL COMPANY NAME PER [[MOS:FIRSTCORP]]. -->,<ref>{{Cite web|title=Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. |url=https://www.bloomberg.com/profile/company/2544189D:CH|website=[[Bloomberg L.P.]]}}</ref><ref>{{Citation|date=19 October 2023 |title=DeepSeek Coder Model Service Agreement |url=https://chat.deepseek.com/downloads/DeepSeek%20Coder%20Model%20Service%20Agreement_1019.pdf|website=DeepSeek}}</ref><ref>{{Cite web |title=DeepSeek Coder Privacy Policy |url=https://chat.deepseek.com/downloads/DeepSeek%20Coder%20Privacy%20Policy_1019.pdf |access-date=2025-02-19 |website=DeepSeek}}</ref>{{Efn|{{Lang-zh|c=杭州深度求索人工智能基础技术研究有限公司}}.<ref>{{Cite web|title=全国互联网安全管理平台|url=https://beian.mps.gov.cn/#/query/webSearch?code=33010502011812|website=beian.mps.gov.cn|access-date=9 February 2025}}</ref> Sometimes simply referred to in English as '''Hangzhou DeepSeek Artificial Intelligence'''.}} [[Trade name|doing business as]] '''DeepSeek''',{{Efn|{{Lang-zh|p=Shēndù Qiúsuǒ|s=深度求索}}}} is a Chinese [[artificial intelligence]] company that develops [[large language model]]s (LLMs). Based in [[Hangzhou, Zhejiang]], it is owned and funded by the Chinese hedge fund [[High-Flyer]]. DeepSeek was founded in July 2023 by High-Flyer co-founder [[Liang Wenfeng]], who also serves as the [[Chief executive officer|CEO]] for both companies. The company launched [[DeepSeek (chatbot)|an eponymous chatbot]] alongside its DeepSeek-R1 model in January 2025.

Released under the [[MIT License]], DeepSeek-R1 provides responses comparable to other contemporary large language models, such as [[OpenAI]]'s [[GPT-4o]] and [[OpenAI o1|o1]].<ref>{{Cite journal |last=Gibney|first=Elizabeth|date=23 January 2025|title=China's cheap, open AI model DeepSeek thrills scientists |url=https://www.nature.com/articles/d41586-025-00229-6|journal=[[Nature (journal)|Nature]]|volume=638 |issue=8049 |pages=13–14 |language=en|doi=10.1038/d41586-025-00229-6|pmid=39849139|bibcode=2025Natur.638...13G |issn=1476-4687}}</ref> Its training cost is reported to be significantly lower than other LLMs. The company claims that it trained its V3 model for US$6&nbsp;million compared to $100&nbsp;million for OpenAI's [[GPT-4]] in 2023,<ref name="vincent">{{cite news|last=Vincent|first=James |date=28 January 2025|title=The DeepSeek panic reveals an AI world ready to blow |url=https://www.theguardian.com/commentisfree/2025/jan/28/deepseek-r1-ai-world-chinese-chatbot-tech-world-western|newspaper=[[The Guardian]]}}</ref> and approximately one-tenth of the computing power used for [[Meta Platforms|Meta]]'s comparable model, [[Llama 3.1]].<ref name="vincent" /><ref name="Metz-2025a">{{Cite news|last1=Metz |first1=Cade|last2=Tobin|first2=Meaghan|date=23 January 2025|title=How Chinese A.I. Start-Up DeepSeek Is Competing With Silicon Valley Giants |url=https://www.nytimes.com/2025/01/23/technology/deepseek-bd-ai-chips.html?smid=fb-nytimes&smtyp=cur&fbclid=IwY2xjawIEynFleHRuA2FlbQIxMQABHZYKXN7GJpUyNRsaGEDQVadxRBarp-aBp1GhiuRe3B57Ehe6HYv7oiK78Q_aem_KTeDgqjV_-R80owNNWOBCQ|work=[[The New York Times]]|language=en-US|issn=0362-4331|access-date=27 January 2025}}</ref><ref>{{cite web|last=Cosgrove |first=Emma|date=27 January 2025|title=DeepSeek's cheaper models and weaker chips call into question trillions in AI infrastructure spending|url=https://www.businessinsider.com/explaining-deepseek-chinese-models-efficiency-scaring-markets-2025-1|website=[[Business Insider]]}}</ref><ref name="Erdil">{{Cite web|last=Erdil|first=Ege|date=17 January 2025|title=How has DeepSeek improved the Transformer architecture? |url=https://epoch.ai/gradient-updates/how-has-deepseek-improved-the-transformer-architecture|access-date=3 February 2025|website=Epoch AI |language=en}}</ref> DeepSeek's success against larger and more established rivals has been described as "upending AI".<ref name="Metz-2025b">{{Cite news|last=Metz|first=Cade|date=27 January 2025|title=What is DeepSeek? And How Is It Upending A.I.? |url=https://www.nytimes.com/2025/01/27/technology/what-is-deepseek-china-ai.html|work=[[The New York Times]]|language=en-US|issn=0362-4331 |access-date=27 January 2025}}</ref><ref>{{Cite news|last=Roose |first=Kevin|date=28 January 2025|title=Why DeepSeek Could Change What Silicon Valley Believe About A.I. |url=https://www.nytimes.com/2025/01/28/technology/why-deepseek-could-change-what-silicon-valley-believes-about-ai.html|access-date=28 January 2025|work=[[The New York Times]]|language=en-US|issn=0362-4331}}</ref>

DeepSeek's models are "open weight", which provides less freedom for modification than true [[open-source software]].<ref name="Delbert">{{cite web |last=Delbert|first=Caroline|date=31 January 2025|title=DeepSeek Is Cracking the 'Black Box' of Corporate AI Wide Open |url=https://www.popularmechanics.com/science/a63633889/deepseek-open-weight/|access-date=12 February 2025|website=[[Popular Mechanics]]}}</ref><ref>{{cite journal|last=Gibney|first=Elizabeth|date=23 January 2025 |title=China's cheap, open AI model DeepSeek thrills scientists|url=https://www.nature.com/articles/d41586-025-00229-6|access-date=12 February 2025|journal=[[Nature (journal)|Nature]]|volume=638 |issue=8049 |pages=13–14 |doi=10.1038/d41586-025-00229-6 |pmid=39849139 |bibcode=2025Natur.638...13G }}</ref> The company reportedly recruits AI researchers from top Chinese universities<ref name="Metz-2025b" /> and hires from outside the [[computer science]] field to diversify its models' knowledge and abilities.<ref name="Metz-2025a" />

The low cost of training and running the language model was attributed to Chinese firms' lack of access to [[Nvidia]] chipsets, which were restricted by the US as part of the ongoing [[trade war]] between the two countries. This breakthrough in reducing expenses while increasing efficiency and maintaining the model's performance power and quality in the AI industry sent "shockwaves" through the market. It threatened the dominance of AI leaders like Nvidia and contributed to the largest drop in US stock market history, with Nvidia alone losing $600 billion in market value.<ref>{{cite web|last=Saah|first=Jasper|date=13 February 2025 |title=DeepSeek sends shock waves across Silicon Valley |url=https://liberationnews.org/deepseek-sends-shock-waves-across-silicon-valley/|access-date=13 February 2025|website=[[Party for Socialism and Liberation|Liberation News – The Newspaper of the Party for Socialism and Liberation]]}}</ref><ref>{{cite web|last=Sillars |first=James|date=28 January 2025|title=DeepSeek: Tech firm suffers biggest drop in US stock market history as low-cost Chinese AI company bites Silicon Valley|url=https://news.sky.com/story/deepseek-us-tech-stocks-tumble-on-fears-of-cheaper-chinese-ai-13297788|access-date=13 February 2025|website=[[Sky News]]}}</ref>

==History==
===Founding and early years (2016–2023)===
In February 2016, High-Flyer was co-founded by AI enthusiast Liang Wenfeng, who had been trading since the [[2007–2008 financial crisis]] while attending [[Zhejiang University]].<ref>{{Cite web|last=Chen |first=Caiwei|date=24 January 2025|title=How a top Chinese AI model overcame US sanctions |url=https://www.technologyreview.com/2025/01/24/1110526/china-deepseek-top-ai-despite-sanctions/|url-status=live|archive-url=https://web.archive.org/web/20250125180427/https://www.technologyreview.com/2025/01/24/1110526/china-deepseek-top-ai-despite-sanctions/|archive-date=25 January 2025|access-date=25 January 2025 |website=[[MIT Technology Review]]|language=en}}</ref>

The company began stock-trading using a [[GPU]]-dependent deep learning model on 21 October 2016. Prior to this, they used [[CPU]]-based models, mainly linear models. Most trading was driven by AI by the end of 2017.<ref name="HI">{{Cite web|title=幻方 {{!}} 幻方历程 |url=https://www.high-flyer.cn/history/|access-date=2025-02-02|website=[[High-Flyer]]|language=zh-CN}}</ref>

In 2019, Liang established High-Flyer as a hedge fund focused on developing and using AI trading algorithms. By 2021, High-Flyer exclusively used AI in trading,<ref name="Ottinger-2024">{{Cite web |last=Ottinger|first=Lily|date=9 December 2024|title=Deepseek: From Hedge Fund to Frontier Model Maker |url=https://www.chinatalk.media/p/deepseek-from-hedge-fund-to-frontier|access-date=28 December 2024|website=ChinaTalk|archive-date=28 December 2024|language=en|archive-url=https://web.archive.org/web/20241228030725/https://www.chinatalk.media/p/deepseek-from-hedge-fund-to-frontier|url-status=live}}</ref> often using [[Nvidia]] chips.<ref name="FT 2025">{{cite news |last1=Olcott|first1=Eleanor|last2=Wu|first2=Zijing|date=24 January 2025|title=How small Chinese AI start-up DeepSeek shocked Silicon Valley|url=https://www.removepaywall.com/search?url=https://www.ft.com/content/747a7b11-dcba-4aa5-8d25-403f56216d7e |access-date=31 January 2025|work=[[Financial Times]]}}</ref>

Initial computing cluster Fire-Flyer began construction in 2019 and finished in 2020, at a cost of 200 million yuan. It contained 1,100 GPUs interconnected at a rate of 200 Gbit/s. It was 'retired' after 1.5 years in operation.<ref name="HI" />

In 2021, Liang began stockpiling Nvidia GPUs for an AI project.<ref name="FT 2025" /> According to [[36Kr]], Liang acquired 10,000 [[Ampere (microarchitecture)#A100 accelerator and DGX A100|Nvidia A100]] GPUs<ref name="CNBC 2023">{{cite news|last=Leswing|first=Kif|date=23 February 2023|title=Meet the $10,000 Nvidia chip powering the race for A.I.|url=https://www.cnbc.com/2023/02/23/nvidias-a100-is-the-10000-chip-powering-the-race-for-ai-.html|work=[[CNBC]]|access-date=30 January 2025}}</ref> before the United States restricted chip sales to China.<ref name="Ottinger-2024" /> Computing cluster Fire-Flyer 2 began construction in 2021 with a budget of 1 billion yuan.<ref name="HI" />

It was reported that in 2022, Fire-Flyer 2's capacity had been used at over 96%, totaling 56.74 million GPU hours. 27% was used to support scientific computing outside the company.<ref name="HI" />

During 2022, Fire-Flyer 2 had 5000 [[PCI Express|PCIe]] A100 GPUs in 625 nodes, each containing 8 GPUs. At the time, they exclusively used PCIe instead of the [[Nvidia DGX|DGX]] version of A100, since at the time the models they trained could fit within a single 40 GB GPU [[Video random-access memory|VRAM]], so there was no need for the higher bandwidth of DGX (i.e. they required only data parallelism but not model parallelism).<ref name="RD" /> Later, they incorporated [[NVLink]]s and NCCL, to train larger models that required model parallelism.<ref name="Deng, Chengqi-2024" /><ref name="DL" />

On 14 April 2023,<ref>{{Cite web|title=独家{{!}}幻方量化回应市场关注：AGI不是用来炒股的，"和金融没关系" |url=https://www.yicai.com/news/101732215.html|access-date=2025-02-03 |website=[[China Business Network|Yicai]]}}</ref> High-Flyer announced the start of an [[artificial general intelligence]] lab dedicated to research developing AI tools separate from High-Flyer's financial business.<ref>{{Cite web|last=Yu|first=Xu|date=17 April 2023|title=[Exclusive] Chinese Quant Hedge Fund High-Flyer Won't Use AGI to Trade Stocks, MD Says |url=https://www.yicaiglobal.com/news/exclusive-chinese-quant-fund-high-flyer-will-not-use-agi-to-trade-stocks-managing-director-says |access-date=28 December 2024|website=[[China Business Network|Yicai Global]]|language=en|archive-date=31 December 2023|archive-url=https://web.archive.org/web/20231231030712/https://www.yicaiglobal.com/news/exclusive-chinese-quant-fund-high-flyer-will-not-use-agi-to-trade-stocks-managing-director-says|url-status=live}}</ref><ref name="scmp 1 January 2025">{{cite web|last1=Jiang|first1=Ben |last2=Perezi|first2=Bien|date=1 January 2025|title=Meet DeepSeek: the Chinese start-up that is changing how AI models are trained |url=https://www.scmp.com/tech/tech-trends/article/3293050/meet-deepseek-chinese-start-changing-how-ai-models-are-trained|work=[[South China Morning Post]]|language=en|access-date=1 January 2025 |archive-date=22 January 2025|archive-url=https://web.archive.org/web/20250122160046/https://www.scmp.com/tech/tech-trends/article/3293050/meet-deepseek-chinese-start-changing-how-ai-models-are-trained|url-status=live}}</ref> Incorporated on 17 July 2023,<ref name="DeepSeek突传消息!" /> with High-Flyer as the investor and backer, the lab became its own company, DeepSeek.<ref name="Ottinger-2024" /><ref name="McMorrow-2024">{{Cite news|last1=McMorrow|first1=Ryan |last2=Olcott|first2=Eleanor|date=9 June 2024|title=The Chinese quant fund-turned-AI pioneer|url=https://www.ft.com/content/357f3c68-b866-4c2e-b678-0d075051a260|url-status=live|archive-url=https://web.archive.org/web/20240717030903/https://www.ft.com/content/357f3c68-b866-4c2e-b678-0d075051a260|archive-date=17 July 2024 |access-date=28 December 2024|work=[[Financial Times]]}}</ref><ref name="scmp 1 January 2025" /> [[Venture capital]] firms were reluctant to provide funding, as they considered it unlikely that the venture would be able to quickly generate an "[[Exit (investing)|exit]]".<ref name="Ottinger-2024" />

On 16 May 2023, the company Beijing DeepSeek Artificial Intelligence Basic Technology Research Company, Limited. was incorporated. It was later taken under 100% control of Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd, which was incorporated 2 months after.{{Citation needed|date=February 2025}}

===Model releases (2023–present)===
On 2 November 2023, DeepSeek released its first model, DeepSeek Coder. On 29 November 2023, DeepSeek released the DeepSeek-LLM series of models.<ref name="Dong, Kai-2024" />{{Pg|location=section 5}} On 9 January 2024, they released 2 DeepSeek-MoE models (Base and Chat).<ref name="Dai-2024" /> In April 2024, they released 3 DeepSeek-Math models: Base, Instruct, and RL.<ref name="PL">{{Citation|last1=Shao |first1=Zhihong|last2=Wang|first2=Peiyi|last3=Zhu|first3=Qihao|last4=Xu|first4=Runxin|last5=Song|first5=Junxiao|last6=Bi|first6=Xiao|last7=Zhang|first7=Haowei|last8=Zhang|first8=Mingchuan|last9=Li|first9=Y. K.|date=27 April 2024|title=DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models|arxiv=2402.03300}}.</ref>

DeepSeek-V2 was released in May 2024. In June 2024, the DeepSeek-Coder V2 series was released.<ref name="V2" />
[[File:Deepseek login error.png|thumb|The DeepSeek login page shortly after a [[cyberattack]] that occurred following its January 20 launch]]
DeepSeek V2.5 was released in September and updated in December 2024.<ref name="HF">{{Cite web|date=3 January 2025|title=deepseek-ai/DeepSeek-V2.5 · Hugging Face|url=https://huggingface.co/deepseek-ai/DeepSeek-V2.5|access-date=28 January 2025|website=[[Hugging Face]]}}</ref> On 20 November 2024, DeepSeek-R1-Lite-Preview became accessible via API and chat.<ref name="DSLI_1" /><ref name="RP" /> In December 2024, the company released the base model DeepSeek-V3-Base and the chat model DeepSeek-V3.<ref name="Deng, Chengqi-2024" />

On 20 January 2025, DeepSeek released the [[DeepSeek (chatbot)|DeepSeek chatbot]], based on the DeepSeek-R1 model, free of charge for [[iOS]] and [[Android (operating system)|Android]]; by 27 January, DeepSeek had surpassed [[ChatGPT]] as the most downloaded freeware app on the [[App Store (Apple)|iOS App Store]] in the United States,<ref name="Metz-2025b" /> causing Nvidia's share price to drop by 18%.<ref>{{Cite web |last=Field|first=Hayden|date=27 January 2025|title=China's DeepSeek AI dethrones ChatGPT on App Store: Here's what you should know |url=https://www.cnbc.com/2025/01/27/chinas-deepseek-ai-tops-chatgpt-app-store-what-you-should-know.html|work=[[CNBC]]}}</ref><ref>{{Cite web|last=Picchi|first=Aimee|date=27 January 2025|title=What is DeepSeek, and why is it causing Nvidia and other stocks to slump?|url=https://www.cbsnews.com/news/what-is-deepseek-ai-china-stock-nvidia-nvda-asml/|work=[[CBS News]]}}</ref>

==Company operation==
Based in [[Hangzhou, Zhejiang]], DeepSeek is owned and funded by the Chinese hedge fund [[High-Flyer]] co-founder [[Liang Wenfeng]], who also serves as its [[Chief executive officer|CEO]].

As of May 2024, Liang owned 84% of DeepSeek through two [[shell corporation]]s.{{NoteTag|宁波程信柔兆企业管理咨询合伙企业（有限合伙） and 宁波程恩企业管理咨询合伙企业（有限合伙）}}<ref>{{Cite web|date=10 May 2024 |title=大模型价格又砍一刀 这次"屠夫"竟是量化私募？ |url=https://www.cls.cn/detail/1672635|access-date=3 February 2025 |website=www.cls.cn}}</ref>

===Strategy===
DeepSeek is focused on research and has not detailed plans for commercialization.<ref name="Schneider-2024">{{Cite web |last=Schneider|first=Jordan|date=27 November 2024|title=Deepseek: The Quiet Giant Leading China's AI Race |url=https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas |website=ChinaTalk|language=en|access-date=28 December 2024}}</ref>

This allows its technology to avoid the most stringent provisions of China's AI regulations, such as requiring consumer-facing technology to comply with government controls on information.<ref name="Metz-2025a" />

DeepSeek's hiring preferences target technical abilities rather than work experience; most new hires are either recent university graduates or developers whose AI careers are less established.<ref name="scmp 1 January 2025" /><ref name="Metz-2025a" />

Likewise, the company recruits individuals without any computer science background to help its technology understand more knowledge areas,<ref name="Metz-2025b" /> such as poetry and China's notoriously difficult [[Gaokao|college admissions exams (Gaokao)]].<ref name="Metz-2025a" />

==Training framework==
High-Flyer/DeepSeek operates at least two computing clusters, Fire-Flyer (萤火一号) and Fire-Flyer 2 (萤火二号). Fire-Flyer 2 consists of co-designed software and hardware architecture. On the hardware side, Nvidia GPUs use 200 [[Data-rate units|Gbps]] interconnects. The cluster is divided into two "zones", and the platform supports cross-zone tasks. The network topology was two [[fat tree]]s, chosen for high [[bisection bandwidth]]. On the software side are:<ref name="DL">{{Cite book|last1=An|first1=Wei|last2=Bi|first2=Xiao |last3=Chen|first3=Guanting|last4=Chen|first4=Shanhuang|last5=Deng|first5=Chengqi|last6=Ding|first6=Honghui|last7=Dong|first7=Kai|last8=Du|first8=Qiushi|last9=Gao|first9=Wenjun|last10=Guan|first10=Kang|last11=Guo|first11=Jianzhong|last12=Guo|first12=Yongqiang|last13=Fu|first13=Zhe|last14=He|first14=Ying|last15=Huang|first15=Panpan |chapter=Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning |date=17 November 2024|title=SC24: International Conference for High Performance Computing, Networking, Storage and Analysis |chapter-url=https://ieeexplore.ieee.org/document/10793193|publisher=IEEE|pages=1–23|doi=10.1109/SC41406.2024.00089|isbn=979-8-3503-5291-7|arxiv=2408.14158}}</ref><ref name="HI" />

*<code>3FS</code> (Fire-Flyer File System): A [[Clustered file system|distributed parallel file system]], specifically designed for asynchronous random reads. It uses Direct I/O and [[Remote direct memory access|RDMA Read]]. In contrast to standard Buffered I/O, Direct I/O does not cache data. Caching is useless for this case, since each data read is random, and isn't reused.<ref>{{Cite web|date=June 13, 2019|title=幻方力量 {{!}} 高速文件系统 3FS|url=https://www.high-flyer.cn/blog/3fs/|access-date=2025-02-03|website=[[High-Flyer]]|language=en}}</ref><ref>{{Citation |title=deepseek-ai/3FS |date=2025-02-28 |url=https://github.com/deepseek-ai/3FS |access-date=2025-02-28 |publisher=DeepSeek}}</ref>
*<code>hfreduce</code>: Library for asynchronous communication, originally designed to replace Nvidia Collective Communication Library (NCCL).<ref name="RD">{{Cite web|date=March 4, 2020 |title=hfreduce {{!}} 高性能的多卡并行通信工具|url=https://www.high-flyer.cn/blog/hf-reduce/|access-date=2025-02-03|website=[[High-Flyer]]|language=en}}</ref> It is mainly used for [[allreduce]], especially of gradients during [[backpropagation]]. It is asynchronously run on the CPU to avoid blocking [[Compute kernel|kernels]] on the GPU.<ref name="DL" /> It uses [[two-tree broadcast]] like NCCL.<ref name="RD" />
*<code>hfai.nn</code>: Software library of commonly used operators for neural network training, similar to <code>torch.nn</code> in [[PyTorch]]. 
*<code>HaiScale Distributed Data Parallel</code> (DDP): Parallel training library that implements various forms of parallelism such as [[Data parallelism|Data Parallelism]] (DP), [[Pipeline (computing)|Pipeline Parallelism]] (PP), Tensor Parallelism (TP), Experts Parallelism (EP), Fully Sharded Data Parallel (FSDP) and Zero Redundancy Optimizer (ZeRO). It is similar to PyTorch DDP, which uses NCCL on the backend.
*<code>HAI Platform</code>: Various applications such as task scheduling, fault handling, and disaster recovery.<ref>{{Citation |date=February 2, 2025|title=HFAiLab/hai-platform |url=https://github.com/HFAiLab/hai-platform|access-date=2025-02-03 |website=[[High-Flyer]]}}</ref>
As of 2022, Fire-Flyer 2 had 5000 [[PCI Express|PCIe]] A100 GPUs in 625 nodes, each containing 8 GPUs.<ref name="RD" /> They later incorporated NVLinks and NCCL, to train larger models that required model parallelism.<ref name="Deng, Chengqi-2024" /><ref name="DL" />

==Development and release history==
{| class="wikitable"
|+Major versions of DeepSeek models. SFT stands for supervised finetuning.
!Major versions
!Release date
!Major variants
!Remarks
|-
|DeepSeek Coder
|2 Nov 2023
|Base (pretrained); Instruct (with instruction-finetuned)
| rowspan="2" |The architecture is essentially the same as Llama.
|-
|DeepSeek-LLM
|29 Nov 2023
|Base;

Chat (with SFT)
|-
|DeepSeek-MoE
|9 Jan 2024
|Base;

Chat
|Developed a variant of [[mixture of experts]] (MoE).
|-
| rowspan="3" |DeepSeek-Math
| rowspan="3" |Apr 2024
|Base
|Initialized with DS-Coder-Base-v1.5
|-
|Instruct (with SFT)
|
|-
|RL (using a process reward model)
|Developed [[Group Relative Policy Optimization|Group Relative Policy Optimization (GRPO)]], a variant of [[Proximal Policy Optimization]] (PPO).
|-
|DeepSeek V2
|May 2024
|DeepSeek-V2, DeepSeek-V2-Chat

DeepSeek-V2-Lite, DeepSeek-V2-Lite-Chat

DeepSeek-Coder-V2

DeepSeek-V2.5
|Developed multi-head latent attention (MLA). Also used mixture of experts (MoE).
|-
|DeepSeek V3
|Dec 2024
|DeepSeek-V3-Base

DeepSeek-V3 (a chat model)
|The architecture is essentially the same as V2.
|-
| rowspan="3" |DeepSeek R1
|20 Nov 2024
|DeepSeek-R1-Lite-Preview
|Only accessed through API and a chat interface.
|-
| rowspan="2" |20 Jan 2025
|DeepSeek-R1

DeepSeek-R1-Zero
|Initialized from DeepSeek-V3-Base and sharing the V3 architecture.
|-
|Distilled models
|Initialized from other models, such as Llama, Qwen, etc. Distilled from data synthesized by R1 and R1-Zero.<ref name="Ma, Shirong-2025" />
|}
The first DeepSeek models were essentially the same as Llama,<ref name="Dong, Kai-2024" /> which were dense decoder-only [[Transformer (deep learning architecture)|Transformers]]. Later models incorporated Mixture of Experts, and then multi-head latent attention.<ref name="Dai-2024" /><ref name="V2" />

A decoder-only Transformer consists of multiple identical decoder layers. Each of these layers features two main components: an attention layer and a FeedForward network (FFN) layer.<ref name="V2" /> In the attention layer, the traditional multi-head attention mechanism has been enhanced with multi-head latent attention. This update introduces compressed latent vectors to boost performance and reduce memory usage during inference.{{Citation needed|date=February 2025}}

Meanwhile, the FFN layer adopts a variant of the mixture of experts (MoE) approach, effectively doubling the number of experts compared to standard implementations. It distinguishes between two types of experts: shared experts, which are always active to encapsulate general knowledge, and routed experts, where only a select few are activated to capture specialized information.{{Citation needed|date=February 2025}}

==Overview of models==
{{Technical|section|date=January 2025}}
{{Primary sources|section|date=February 2025}}
DeepSeek's models are "open weight", which provides less freedom for modification than true [[open source]] software.<ref name="Delbert"/><ref>{{cite journal|last=Gibney|first=Elizabeth |date=23 January 2025|title=China's cheap, open AI model DeepSeek thrills scientists|url=https://www.nature.com/articles/d41586-025-00229-6|access-date=12 February 2025|journal=[[Nature (journal)|Nature]]|volume=638 |issue=8049 |pages=13–14 |doi=10.1038/d41586-025-00229-6 |pmid=39849139 |bibcode=2025Natur.638...13G }}</ref>

===DeepSeek Coder===
DeepSeek Coder is a series of 8 models, 4 pretrained (<code>Base</code>) and 4 instruction-finetuned (<code>Instruct</code>). They all have 16K context lengths. The model was made [[source-available]] under the DeepSeek License, which includes "open and responsible downstream usage" restrictions.<ref>{{Cite web |title=DeepSeek-Coder/LICENSE-MODEL at main · deepseek-ai/DeepSeek-Coder|url=https://github.com/deepseek-ai/DeepSeek-Coder/blob/main/LICENSE-MODEL|url-status=live|archive-url=https://web.archive.org/web/20250122195853/https://github.com/deepseek-ai/deepseek-coder/blob/main/LICENSE-MODEL|archive-date=22 January 2025|access-date=24 January 2025|website=[[GitHub]]|language=en}}</ref>

The [[Training, validation, and test data sets|training]] program was:<ref name="Guo-2024">{{Citation|title=DeepSeek-Coder: When the Large Language Model Meets Programming – The Rise of Code Intelligence |date=26 January 2024 |arxiv=2401.14196|last1=Guo|first1=Daya|last2=Zhu|first2=Qihao|last3=Yang|first3=Dejian|last4=Xie|first4=Zhenda|last5=Dong|first5=Kai|last6=Zhang|first6=Wentao|last7=Chen|first7=Guanting|last8=Bi|first8=Xiao |last9=Wu|first9=Y.}}</ref><ref>{{Cite web|title=DeepSeek Coder |url=https://deepseekcoder.github.io/|access-date=27 January 2025 |website=deepseekcoder.github.io}}</ref><ref>{{Citation|date=27 January 2025|title=deepseek-ai/DeepSeek-Coder|url=https://github.com/deepseek-ai/deepseek-coder/|access-date=27 January 2025|publisher=DeepSeek}}</ref>

# Pretraining: 1.8T tokens (87% source code, 10% code-related English (GitHub markdown and [[Stack Exchange]]), and 3% code-unrelated Chinese).
# Long-context pretraining: 200B tokens. This extends the context length from 4K to 16K. This produced the <code>Base</code> models.
# Supervised [[Fine-tuning (deep learning)|finetuning]] (SFT): 2B tokens of instruction data. This produced the <code>Instruct</code> models.

They were trained on clusters of A100 and [[Hopper (microarchitecture)|H800]] Nvidia GPUs, connected by [[InfiniBand]], [[NVLink]], [[NVSwitch]].<ref name="Guo-2024" />

{| class="wikitable sortable"
|+DeepSeek Coder properties<ref name="Guo-2024" />{{Pg|location=Table 2}}<ref>{{Cite web|title=deepseek-ai/deepseek-coder-5.7bmqa-base · Hugging Face|url=https://huggingface.co/deepseek-ai/deepseek-coder-5.7bmqa-base|access-date=27 January 2025|website=[[Hugging Face]]}}</ref>
|-
! {{abbr|Params|Parameters}}. !! <math>n_{\text{layers}}</math> !! <math>d_{\text{model}}</math> 
!<math>d_{\text{intermediate}}</math>!! <math>n_{\text{heads}}</math> !! <math>n_{\text{kv-heads}} </math>
|-
| 1.3B || 24 || 2048
|5504|| 16 ||  16
|-
|5.7B
|32
|4096
|11008
|32
|1{{NoteTag|name=fn1|The number of heads does not equal the number of KV heads, due to GQA.}}
|-
| 6.7B || 32 || 4096 
|11008|| 32 || 32 
|-
| 33B || 62 || 7168
|19200|| 56 ||7{{NoteTag|name=fn1}}
|}

===DeepSeek-LLM===
The DeepSeek-LLM series was released in November 2023. It has 7B and 67B parameters in both Base and Chat forms. DeepSeek's accompanying paper claimed benchmark results higher than [[Llama 2]] and most open-source LLMs at the time.<ref name="Dong, Kai-2024">{{Citation |author1=DeepSeek-AI|arxiv=2401.02954|last2=Bi|first2=Xiao|last3=Chen |first3=Deli|last4=Chen|first4=Guanting|last5=Chen|first5=Shanhuang|last6=Dai|first6=Damai|last7=Deng|first7=Chengqi|last8=Ding|first8=Honghui|last9=Dong|first9=Kai|date=5 January 2024|title=DeepSeek LLM: Scaling Open-Source Language Models with Longtermism}}</ref>{{Pg|location=section 5}} The model code is under the source-available DeepSeek License.<ref>{{Citation|date=27 January 2025|title=deepseek-ai/DeepSeek-LLM|url=https://github.com/deepseek-ai/DeepSeek-LLM|access-date=27 January 2025|publisher=DeepSeek}}</ref>

The architecture was essentially the same as the [[Llama (language model)|Llama]] series. They used the [[Transformer (deep learning architecture)#pre-LN|pre-norm]] [[Transformer (deep learning architecture)#decoder-only|decoder-only Transformer]] with [[RMSNorm]] as the normalization, [[SwiGLU]] in the feedforward layers, [[rotary positional embedding]] (RoPE), and [[grouped-query attention]] (GQA). Both had vocabulary size 102,400 ([[Byte pair encoding#Byte-level BPE|byte-level BPE]]) and context length of 4096. They trained on 2&nbsp;trillion tokens of English and Chinese text obtained by deduplicating the [[Common Crawl]].<ref name="Dong, Kai-2024" />

{| class="wikitable sortable"
|+DeepSeek LLM properties<ref name="Dong, Kai-2024" />{{Pg|location=Table 2}}
|-
! {{abbr|Params|Parameters}}. !! <math>n_{\text{layers}}</math> !! <math>d_{\text{model}}</math> 
!<math>d_{\text{intermediate}}</math>!! <math>n_{\text{heads}}</math> !! <math>n_{\text{kv-heads}} </math>
|-
| 7B || 30 || 4096 
|11008|| 32 || 32 
|-
| 67B || 95 || 8192 
|22016|| 64 || 8{{NoteTag|name=fn1}}
|}
The Chat versions of the two Base models was released concurrently, obtained by training Base by [[Reinforcement learning from human feedback|supervised finetuning (SFT) followed by direct policy optimization (DPO)]].<ref name="Dong, Kai-2024" />

====MoE====
DeepSeek-MoE models (Base and Chat), each have 16B parameters (2.7B activated per token, 4K context length). The training was essentially the same as DeepSeek-LLM 7B, and was trained on a part of its training dataset. They claimed performance comparable to a 16B MoE as a 7B non-MoE. It is a variant of the standard [[Mixture of experts#Sparsely-gated MoE layer|sparsely-gated MoE]], with "shared experts" that are always queried, and "routed experts" that might not be. They found this to help with expert balancing. In standard MoE, some experts can become overused, while others are rarely used, wasting space. Attempting to balance expert usage causes experts to replicate the same capacity. They proposed the shared experts to learn core capacities that are often used, and let the routed experts learn peripheral capacities that are rarely used.<ref name="Dai-2024">{{Citation|last1=Dai|first1=Damai|last2=Deng|first2=Chengqi |last3=Zhao|first3=Chenggang|last4=Xu|first4=R. X.|last5=Gao |first5=Huazuo|last6=Chen|first6=Deli|last7=Li|first7=Jiashi|last8=Zeng|first8=Wangding|last9=Yu|first9=Xingkai|date=11 January 2024 |title=DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models|arxiv=2401.06066}}</ref>

====Math====
DeepSeek-Math includes 3 models: Base, Instruct, and RL. Math was trained as follows:<ref name="PL" />

# Initialize with a previously pretrained DeepSeek-Coder Base v1.5 7B.
# Further pretrain with 500B tokens (6% DeepSeekMath Corpus, 4% AlgebraicStack, 10% arXiv, 20% GitHub code, 10% Common Crawl). This produced Base.
# Train an instruction-following model by SFT Base with 776K math problems and tool-use-integrated step-by-step solutions. This produced Instruct.
# [[Reinforcement learning]] (RL): The reward model was a [[Reasoning language model#PRM|process reward model]] (PRM) trained from Base according to the Math-Shepherd method.<ref>{{Citation|last1=Wang|first1=Peiyi|last2=Li|first2=Lei|last3=Shao|first3=Zhihong|last4=Xu|first4=R. X.|last5=Dai|first5=Damai|last6=Li |first6=Yifei|last7=Chen|first7=Deli|last8=Wu|first8=Y.|last9=Sui|first9=Zhifang|date=19 February 2024|title=Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations |arxiv=2312.08935}}.</ref> This reward model was then used to train Instruct using [[Group Relative Policy Optimization]] (GRPO) on a dataset of 144K math questions "related to GSM8K and MATH". The reward model was continuously updated during training to avoid reward hacking. This resulted in RL.

===V2===
[[File:DeepSeek_MoE_and_MLA_(DeepSeek-V2).svg|thumb|upright=1.4|The architecture of V2, showing both shared-routed MoE and MLA<ref name="Ruan, Chong-2024" />{{Pg|location=Figure 2}}]]
In May 2024, DeepSeek released the DeepSeek-V2 series. The series includes 4 models, 2 base models (DeepSeek-V2, DeepSeek-V2 Lite) and 2 chatbots (Chat). The two larger models were trained as follows:<ref name="Ruan, Chong-2024">{{Citation|author1=DeepSeek-AI|last2=Liu|first2=Aixin|last3=Feng|first3=Bei|last4=Wang|first4=Bin|last5=Wang|first5=Bingxuan|last6=Liu|first6=Bo|last7=Zhao|first7=Chenggang|last8=Dengr|first8=Chengqi|last9=Ruan|first9=Chong|date=19 June 2024|title=DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model|arxiv=2405.04434}}.</ref>

# Pretrain on a dataset of 8.1T tokens, using 12% more Chinese tokens than English ones.
# Extend context length from 4K to 128K using YaRN.<ref name="Peng-2023">{{Citation|last1=Peng|first1=Bowen|last2=Quesnelle|first2=Jeffrey|last3=Fan|first3=Honglu|last4=Shippole|first4=Enrico|date=1 November 2023|title=YaRN: Efficient Context Window Extension of Large Language Models|arxiv=2309.00071}}.</ref> This resulted in DeepSeek-V2.
# SFT with 1.2M instances for helpfulness and 0.3M for safety. This resulted in Chat SFT, which was not released.
# RL using GRPO in two stages. The first stage was trained to solve math and coding problems. This stage used 1 reward model, trained on compiler feedback (for coding) and ground-truth labels (for math). The second stage was trained to be helpful, safe, and follow rules. This stage used 3 reward models. The helpfulness and safety reward models were trained on human preference data. The rule-based reward model was manually programmed. All trained reward models were initialized from Chat (SFT). This resulted in the released version of Chat.

They opted for 2-staged RL, because they found that RL on reasoning data had "unique characteristics" different from RL on general data. For example, RL on reasoning could improve over more training steps.<ref name="Ruan, Chong-2024" />

The two V2-Lite models were smaller, and trained similarly. DeepSeek-V2 Lite-Chat underwent only SFT, not RL. They trained the Lite version to help "further research and development on MLA and DeepSeekMoE".<ref name="Ruan, Chong-2024" />

Architecturally, the V2 models were significantly different from the DeepSeek LLM series. They changed the standard attention mechanism by a [[low-rank approximation]] called [[Transformer (deep learning architecture)#MLA|multi-head latent attention]] (MLA), and used the previously published [[mixture of experts]] (MoE) variant.<ref name="Dai-2024" />

{| class="wikitable sortable"
|+DeepSeek V2 properties<ref name="Ruan, Chong-2024" />{{Pg|location=Section 3.1.2, Appendix B}}<ref>{{Cite web|date=15 May 2024|title=config.json · deepseek-ai/DeepSeek-V2-Lite at main |url=https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/blob/main/config.json|access-date=28 January 2025|website=[[Hugging Face]]}}</ref><ref>{{Cite web|date=6 May 2024 |title=config.json · deepseek-ai/DeepSeek-V2 at main |url=https://huggingface.co/deepseek-ai/DeepSeek-V2/blob/main/config.json|access-date=28 January 2025|website=[[Hugging Face]]}}</ref>
|-
!Name
! {{abbr|Params|Parameters}}. 
!Active {{abbr|params|Parameters}}!! <math>n_{\text{layers}}</math>!! Context length 
!<math>n_{\text{shared experts}}</math>!! <math>n_{\text{routed experts}}</math>
|-
|V2-Lite
| 15.7B 
|2.4B|| 27 || 32K 
|2|| 64 
|-
|V2
| 236B 
|21B|| 60 || 128K 
|2|| 160 
|}
The ''[[Financial Times]]'' reported that it was cheaper than its peers with a price of 2 [[Renminbi|RMB]] for every million output tokens. The [[University of Waterloo]] Tiger Lab's leaderboard ranked DeepSeek-V2 seventh on its LLM ranking.<ref name="McMorrow-2024" />

The DeepSeek-Coder V2 series included V2-Base, V2-Lite-Base, V2-Instruct, and V20-Lite-Instruct.. Training:<ref name="V2">{{Citation|author1=DeepSeek-AI|last2=Zhu|first2=Qihao|last3=Guo |first3=Daya|last4=Shao|first4=Zhihong|last5=Yang|first5=Dejian|last6=Wang|first6=Peiyi|last7=Xu|first7=Runxin|last8=Wu|first8=Y.|last9=Li|first9=Yukun|date=17 June 2024|title=DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence |arxiv=2406.11931}}</ref>{{NoteTag|Inexplicably, the model named <code>DeepSeek-Coder-V2 Chat</code> in the paper was released as <code>DeepSeek-Coder-V2-Instruct</code> in HuggingFace.}}

# Base models were initialized from corresponding intermediate checkpoints after pretraining on 4.2T tokens (not the version at the end of pretraining), then pretrained further for 6T tokens, then context-extended to 128K context length. 
# DeepSeek-Coder and DeepSeek-Math were used to generate 20K code-related and 30K math-related instruction data, then combined with an instruction dataset of 300M tokens. This was used for SFT.
# RL with GRPO. The reward for math problems was computed by comparing with the ground-truth label. The reward for code problems was generated by a reward model trained to predict whether a program would pass the unit tests.

DeepSeek-V2.5 was made by combining DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct.<ref name="HF" />

===V3===
[[File:Multi-Token_Prediction_(DeepSeek)_01.svg|thumb|upright=1.4|Multi-Token Prediction]]
DeepSeek-V3-Base and DeepSeek-V3 (a chat model) use essentially the same architecture as V2 with the addition of [[Transformer (deep learning architecture)#Multi-Token Prediction|multi-token prediction]], which (optionally) decodes extra tokens faster but less accurately. Training:<ref name="Deng, Chengqi-2024">{{Citation|author1=DeepSeek-AI|last2=Liu|first2=Aixin|last3=Feng |first3=Bei|last4=Xue|first4=Bing|last5=Wang|first5=Bingxuan|last6=Wu|first6=Bochao|last7=Lu|first7=Chengda|last8=Zhao|first8=Chenggang|last9=Deng|first9=Chengqi|date=27 December 2024|title=DeepSeek-V3 Technical Report|arxiv=2412.19437}}</ref>

# Pretraining on 14.8T tokens of a multilingual corpus, mostly English and Chinese. It contained a higher ratio of math and programming than the pretraining dataset of V2.
# Extend context length twice, from 4K to 32K and then to 128K, using YaRN.<ref name="Peng-2023" /> This produced DeepSeek-V3-Base.
# SFT for 2 epochs on 1.5M samples of reasoning (math, programming, logic) and non-reasoning (creative writing, roleplay, simple question answering) data. Reasoning data was generated by "expert models". Non-reasoning data was generated by DeepSeek-V2.5 and checked by humans.
#* The "expert models" were trained by starting with an unspecified base model, then SFT on both <problem, original response> data, and synthetic <system prompt, prompt, problem, R1 response> data generated by an internal DeepSeek-R1-Lite model. The system prompt asked R1 to reflect and verify during thinking. Then the expert models were RL using an undisclosed reward function.
#* Each expert model was trained to generate just synthetic reasoning data in one specific domain (math, programming, logic). 
#* Expert models were used instead of R1 itself, since the output from R1 itself suffered "overthinking, poor formatting, and excessive length".
# Model-based reward models were made by starting with a SFT checkpoint of V3, then finetuning on human preference data containing both final reward and chain-of-thought leading to the final reward. The reward model produced reward signals for both questions with objective but free-form answers, and questions without objective answers (such as creative writing).
# An SFT checkpoint of V3 was trained by GRPO using both reward models and rule-based reward. The rule-based reward was computed for math problems with a final answer (put in a box), and for programming problems by unit tests. This produced DeepSeek-V3.

{| class="wikitable sortable"
|+DeepSeek V3 properties<ref name="Deng, Chengqi-2024" />{{Pg|location=Section 4.2}}<ref>{{Cite web|date=26 December 2024|title=config.json · deepseek-ai/DeepSeek-V3 at main|url=https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/config.json|access-date=28 January 2025 |website=[[Hugging Face]]}}</ref>
|-
!Name
! {{abbr|Params|Parameters}}. 
!Active {{abbr|params|Parameters}}!! <math>n_{\text{layers}}</math> !! Context length 
!<math>n_{\text{shared experts}}</math>!! <math>n_{\text{routed experts}}</math> 
|-
|V3
| 671B 
|37B|| 61 || 128K 
|1|| 256 
|}
[[File:Mixed-precision_training_in_DeepSeek_V3.svg|thumb|upright=1.4|Mixed-precision framework for <code>V3</code><ref name="Deng, Chengqi-2024" />{{Pg|location=Figure 6}}]]
The DeepSeek team performed extensive low-level engineering to improve efficiency. They used [[mixed-precision arithmetic]]. Much of the forward pass was performed in [[Floating-point arithmetic|8-bit floating point numbers]] (5E2M: 5-bit exponent and 2-bit [[Mantissa (floating point number)|mantissa]]) rather than the standard [[Single-precision floating-point format|32-bit]], requiring special [[General matrix multiply|GEMM]] routines to accumulate accurately. They used a custom 12-bit float (E5M6) only for the inputs to the linear layers after the attention modules. Optimizer states were in 16-bit ([[Bfloat16 floating-point format|BF16]]). They minimized communication latency by extensively overlapping computation and communication, such as dedicating 20 streaming multiprocessors out of 132 per H800 for only inter-GPU communication. They lowered communication by rearranging (every 10 minutes) the exact machine each expert was on so as to avoid querying certain machines more often than others, adding auxiliary load-balancing losses to the training loss function, and other load-balancing techniques.<ref name="Deng, Chengqi-2024" />

After training, it was deployed on clusters of H800 GPUs. The 8 H800 GPUs within a cluster were connected by NVLink, and the clusters were connected by InfiniBand.<ref name="Deng, Chengqi-2024" />

{| class="wikitable"
|+ Total cost of training the DeepSeek-V3 model<ref name="Deng, Chengqi-2024" />{{Pg|location=Table 1}}
! Stage
! Cost (in one thousand GPU hours)
! Cost (in one million USD$)
|-
| Pre-training
| 2,664
| 5.328
|-
| Context extension
| 119
| 0.24
|-
| Fine-tuning
| 5
| 0.01
|-
| Total
| 2,788
| 5.576
|}

The cost has been discussed<ref>{{Cite web|last1=Patel|first1=Dylan |last2=Kourabi|first2=AJ|last3=O'Laughlin|first3=Dylan|last4=Knuhtsen|first4=Doug|date=31 January 2025|title=DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts |url=https://semianalysis.com/2025/01/31/deepseek-debates/|access-date=13 February 2025|website=SemiAnalysis|language=en-US}}</ref><ref>{{Cite web|last=Thubron|first=Rob|date=3 February 2025|title=DeepSeek's AI costs far exceed $5.5 million claim, may have reached $1.6 billion with 50,000 Nvidia GPUs|url=https://www.techspot.com/news/106612-deepseek-ai-costs-far-exceed-55-million-claim.html|access-date=13 February 2025|website=TechSpot|language=en-US}}</ref><ref>{{Cite web |last=Kajal|first=Kapil|date=31 January 2025|title=Research exposes DeepSeek's AI training cost is not $6M, it's a staggering $1.3B |url=https://www.yahoo.com/news/research-exposes-deepseek-ai-training-165025904.html|access-date=13 February 2025|website=[[Yahoo News]]|language=en-US}}</ref> and called misleading, because it covers only parts of the true cost.<ref>{{Cite web|date=28 January 2025 |title=Martin Vechev of INSAIT: "DeepSeek $6M Cost Of Training Is Misleading"|url=https://therecursive.com/martin-vechev-of-insait-deepseek-6m-cost-of-training-is-misleading/|access-date=13 February 2025|website=TheRecursive.com|language=en-GB}}</ref>

Benchmark tests show that V3 outperformed [[Llama (language model)|Llama]] 3.1 and [[Qwen]] 2.5 while matching [[GPT-4o]] and [[Claude (language model)|Claude]] 3.5 Sonnet.<ref name="scmp 1 January 2025" /><ref>{{Cite web|last=Jiang|first=Ben|date=27 December 2024|title=Chinese start-up DeepSeek's new AI model outperforms Meta, OpenAI products|url=https://www.scmp.com/tech/tech-trends/article/3292507/chinese-start-deepseek-launches-ai-model-outperforms-meta-openai-products|access-date=28 December 2024|work=[[South China Morning Post]]|language=en|archive-date=27 December 2024|archive-url=https://web.archive.org/web/20241227191529/https://www.scmp.com/tech/tech-trends/article/3292507/chinese-start-deepseek-launches-ai-model-outperforms-meta-openai-products|url-status=live}}</ref><ref>{{Cite web|last=Sharma|first=Shubham|date=26 December 2024 |title=DeepSeek-V3, ultra-large open-source AI, outperforms Llama and Qwen on launch|url=https://venturebeat.com/ai/deepseek-v3-ultra-large-open-source-ai-outperforms-llama-and-qwen-on-launch/|access-date=28 December 2024|website=[[VentureBeat]]|language=en-US|archive-date=27 December 2024|archive-url=https://web.archive.org/web/20241227195503/https://venturebeat.com/ai/deepseek-v3-ultra-large-open-source-ai-outperforms-llama-and-qwen-on-launch/|url-status=live}}</ref><ref>{{cite web|last=Wiggers |first=Kyle|date=26 December 2024|title=DeepSeek's new AI model appears to be one of the best 'open' challengers yet |url=https://techcrunch.com/2024/12/26/deepseeks-new-ai-model-appears-to-be-one-of-the-best-open-challengers-yet/|access-date=31 December 2024|website=[[TechCrunch]]|archive-date=2 January 2025 |archive-url=https://web.archive.org/web/20250102103526/https://techcrunch.com/2024/12/26/deepseeks-new-ai-model-appears-to-be-one-of-the-best-open-challengers-yet/|url-status=live}}</ref>

===R1===
{{See also|Reflection (artificial intelligence)}}
In January 2025, DeepSeek released the DeepSeek-R1 model under the [[MIT License]].<ref>{{cite web|last=Edwards|first=Benj|date=21 January 2025|title=Cutting-edge Chinese "reasoning" model rivals OpenAI o1—and it's free to download|url=https://arstechnica.com/ai/2025/01/china-is-catching-up-with-americas-best-reasoning-ai-models/|website=[[Ars Technica]]|access-date=16 February 2025}}</ref>

DeepSeek-R1-Lite-Preview<ref name="DSLI_1">{{cite web|title=Deepseek Log in page |url=https://chat.deepseek.com/sign_in|website=DeepSeek|access-date=30 January 2025}}</ref><ref name="RP">{{Cite web|title=News {{!}} DeepSeek-R1-Lite Release 2024/11/20: 🚀 DeepSeek-R1-Lite-Preview is now live: unleashing supercharged reasoning power!|url=https://api-docs.deepseek.com/news/news1120|archive-url=https://web.archive.org/web/20241120141324/https://api-docs.deepseek.com/news/news1120|archive-date=20 November 2024|access-date=28 January 2025|website=DeepSeek API Docs|language=en}}</ref>{{NoteTag|At that time, the <code>R1-Lite-Preview</code> required selecting "Deep Think enabled", and every user could use it only 50 times a day.}} was trained for logical inference, mathematical reasoning, and real-time problem-solving. DeepSeek claimed that it exceeded performance of [[OpenAI o1]] on benchmarks such as [[American Invitational Mathematics Examination]] (AIME) and MATH.<ref>{{Cite web|last=Franzen|first=Carl|date=20 November 2024 |title=DeepSeek's first reasoning model R1-Lite-Preview turns heads, beating OpenAI o1 performance |url=https://venturebeat.com/ai/deepseeks-first-reasoning-model-r1-lite-preview-turns-heads-beating-openai-o1-performance/|url-status=live|archive-url=https://web.archive.org/web/20241122010413/https://venturebeat.com/ai/deepseeks-first-reasoning-model-r1-lite-preview-turns-heads-beating-openai-o1-performance/|archive-date=22 November 2024|access-date=28 December 2024|website=[[VentureBeat]]|language=en-US}}</ref> However, ''[[The Wall Street Journal]]'' reported that on 15 problems from the 2024 edition of AIME, the o1 model reached a solution faster.<ref>{{Cite web|last=Huang|first=Raffaele|date=24 December 2024|title=Don't Look Now, but China's AI Is Catching Up Fast |url=https://www.wsj.com/tech/ai/china-ai-advances-us-chips-7838fd20 |url-status=live|archive-url=https://web.archive.org/web/20241227183842/https://www.wsj.com/tech/ai/china-ai-advances-us-chips-7838fd20|archive-date=27 December 2024|access-date=28 December 2024|website=[[The Wall Street Journal]] |language=en-US}}</ref>

DeepSeek-R1 and DeepSeek-R1-Zero<ref>{{Cite web|title=Release DeepSeek-R1 · deepseek-ai/DeepSeek-R1@23807ce |url=https://github.com/deepseek-ai/DeepSeek-R1/commit/23807ced51627276434655dd9f27725354818974|access-date=21 January 2025|website=[[GitHub]]|language=en|archive-date=21 January 2025|archive-url=https://web.archive.org/web/20250121104009/https://github.com/deepseek-ai/DeepSeek-R1/commit/23807ced51627276434655dd9f27725354818974 |url-status=live}}</ref> were initialized from DeepSeek-V3-Base and share its architecture. DeepSeek-R1-Distill models were instead initialized from other pretrained open-weight models, including [[Llama (language model)|LLaMA]] and [[Qwen]], then fine-tuned on [[synthetic data]] generated by R1.<ref name="Ma, Shirong-2025">{{Citation|author1=DeepSeek-AI|last2=Guo|first2=Daya|last3=Yang |first3=Dejian|last4=Zhang|first4=Haowei|last5=Song|first5=Junxiao|last6=Zhang|first6=Ruoyu|last7=Xu|first7=Runxin|last8=Zhu|first8=Qihao|last9=Ma|first9=Shirong|date=22 January 2025|title=DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning |arxiv=2501.12948}}</ref>

{{Quote box
| title = Template for <code>DeepSeek-R1-Zero</code>
| quote = A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think> <answer> answer here </answer>. User: <prompt>. Assistant:
| source = – <prompt> is replaced with the specific reasoning question during training.
| align = right
| width = 350px
}}

DeepSeek-R1-Zero was trained exclusively using GRPO RL without SFT. Unlike previous versions, it used no model-based reward. All reward functions were rule-based, "mainly" of two types (other types were not specified): accuracy rewards and format rewards. Accuracy reward was checking whether a boxed answer is correct (for math) or whether a code passes tests (for programming). Format reward was checking whether the model puts its thinking trace within a <think>...</think> tag.<ref name="Ma, Shirong-2025" />

R1-Zero has issues with readability and mixing languages. R1 was trained to address these issues and further improve reasoning:<ref name="Ma, Shirong-2025" />

# SFT DeepSeek-V3-Base on "thousands" of "cold-start" data all with the standard format of <code>|special_token|<reasoning_process>|special_token|<summary></code>, designed to improve model output readability.
# Apply the same GRPO RL process as R1-Zero, adding a "language consistency reward" to encourage it to respond monolingually. This produced an un released internal model.
# Synthesize 600K reasoning data from the internal model, with rejection sampling (i.e. if the generated reasoning had a wrong final answer, then it is removed). Synthesize 200K non-reasoning data (writing, factual QA, self-cognition, translation) using DeepSeek-V3.
# SFT DeepSeek-V3-Base on the 800K synthetic data for 2 epochs.
# Apply the same GRPO RL process as R1-Zero with rule-based reward (for reasoning tasks), but also model-based reward (for non-reasoning tasks, helpfulness, and harmlessness). This produced DeepSeek-R1.

Distilled models were trained by SFT on 800K data synthesized from DeepSeek-R1, in a similar way as step 3. They were not trained with RL.<ref name="Ma, Shirong-2025" />

R2, the successor to R1, is originally planned for release in early May 2025, but release schedule accelerated.<ref>{{cite web |author1=Eduardo Baptista |author2=Julie Zhu |author3=Fanny Potkin |date=February 25, 2025 |title=DeepSeek rushes to launch new AI model as China goes all in |url=https://www.reuters.com/technology/artificial-intelligence/deepseek-rushes-launch-new-ai-model-china-goes-all-2025-02-25/ |access-date=February 25, 2025 |website=Reuters}}</ref>

==Significance==
DeepSeek's success against larger and more established rivals has been described as "upending AI".<ref name="Metz-2025b" /><ref>{{Cite news|last=Roose|first=Kevin|date=28 January 2025|title=Why DeepSeek Could Change What Silicon Valley Believe About A.I. |url=https://www.nytimes.com/2025/01/28/technology/why-deepseek-could-change-what-silicon-valley-believes-about-ai.html|access-date=28 January 2025|work=[[The New York Times]]|language=en-US |issn=0362-4331}}</ref>

The DeepSeek-R1 model provides responses comparable to other contemporary large language models, such as [[OpenAI]]'s [[GPT-4o]] and [[OpenAI o1|o1]].<ref>{{Cite journal|last=Gibney|first=Elizabeth |date=23 January 2025|title=China's cheap, open AI model DeepSeek thrills scientists|url=https://www.nature.com/articles/d41586-025-00229-6|journal=[[Nature (journal)|Nature]]|volume=638 |issue=8049 |pages=13–14 |language=en |doi=10.1038/d41586-025-00229-6|issn=1476-4687|pmid=39849139|bibcode=2025Natur.638...13G }}</ref> Its [[Training, validation, and test data sets|training]] cost is reported to be significantly lower than other LLMs.

The company claims that it trained V3, a predecessor of R1, for US$6&nbsp;million compared to $100&nbsp;million for OpenAI's [[GPT-4]] in 2023,<ref name="vincent" /> and approximately one tenth of the computing power used for [[Meta Platforms|Meta]]'s comparable model, [[Llama (language model)|LLaMA 3.1]].<ref name="vincent" /><ref name="Metz-2025a" /><ref>{{cite web|last=Cosgrove|first=Emma|date=27 January 2025|title=DeepSeek's cheaper models and weaker chips call into question trillions in AI infrastructure spending |url=https://www.businessinsider.com/explaining-deepseek-chinese-models-efficiency-scaring-markets-2025-1|website=[[Business Insider]]}}</ref><ref name="Erdil"/>

Domestically, DeepSeek models offer performance for a low price, and have become the catalyst for China's AI model [[price war]].{{Citation needed|date=February 2025}} It was dubbed the "[[Pinduoduo]] of AI", and other Chinese tech giants such as [[ByteDance]], [[Tencent]], [[Baidu]], and [[Alibaba Group|Alibaba]] cut the price of their AI models. Despite its low price, it was profitable compared to its money-losing rivals.<ref name="Schneider-2024" />

== See also ==
{{Portal|Free and open-source software|Business|China}}
* {{annotated link|DeepSeek (chatbot)}}
* {{annotated link|Artificial intelligence industry in China}}
* {{annotated link|OpenAI}}
* {{annotated link|Jevons paradox}}

== Notes ==
{{notelist}}
{{reflist|group=note}}

==References==
{{reflist|30em|refs=
}}

== External links ==
{{Commons category}}
* {{Official website}}
* [https://github.com/deepseek-ai DeepSeek] on [[GitHub]]
* [https://huggingface.co/deepseek-ai/DeepSeek-V2.5-1210 DeepSeek] on [[Hugging Face]]
* [https://api-docs.deepseek.com/ Official API documentation]
* [https://huggingface.co/collections/Presidentlin/deepseek-papers-674c536aa6acddd9bc98c2ac Anthology of DeepSeek papers]
* [https://www.high-flyer.cn/blog/ Research blog of High-Flyer]

{{Generative AI chatbots}}
{{Generative AI}}
{{Authority control}}

[[Category:2023 establishments in China]]
[[Category:Artificial intelligence companies]]
[[Category:Artificial intelligence laboratories]]
[[Category:Companies based in Hangzhou]]
[[Category:Technology companies established in 2023]]