This Stage used 3 Reward Models

DeepSeek (Chinese: 深度求索; pinyin: Shēndù Qiúsuǒ) is a Chinese synthetic intelligence company that develops open-source big language models (LLMs). Based in Hangzhou, Zhejiang, it is owned and moneyed by Chinese hedge fund High-Flyer, whose co-founder, Liang Wenfeng, established the business in 2023 and acts as its CEO.

The DeepSeek-R1 design offers responses equivalent to other contemporary large language designs, such as OpenAI’s GPT-4o and o1. [1] It is trained at a substantially lower cost-stated at US$ 6 million compared to $100 million for OpenAI’s GPT-4 in 2023 [2] -and needs a tenth of the computing power of a similar LLM. [2] [3] [4] DeepSeek’s AI models were developed in the middle of United States sanctions on India and China for Nvidia chips, [5] which were meant to restrict the ability of these 2 countries to develop innovative AI systems. [6] [7]

On 10 January 2025, DeepSeek released its first free chatbot app, based on the DeepSeek-R1 design, for iOS and Android; by 27 January, DeepSeek-R1 had actually surpassed ChatGPT as the most-downloaded totally free app on the iOS App Store in the United States, [8] triggering Nvidia’s share price to drop by 18%. [9] [10] DeepSeek’s success versus larger and more established rivals has been described as “upending AI”, [8] making up “the very first shot at what is becoming a global AI area race”, [11] and introducing “a brand-new age of AI brinkmanship”. [12]

DeepSeek makes its generative expert system algorithms, designs, and training information open-source, allowing its code to be freely readily available for use, modification, viewing, and developing documents for developing functions. [13] The business apparently vigorously recruits young AI scientists from top Chinese universities, [8] and works with from outside the computer technology field to diversify its models’ understanding and abilities. [3]

In February 2016, High-Flyer was co-founded by AI lover Liang Wenfeng, who had been trading considering that the 2007-2008 monetary crisis while going to Zhejiang University. [14] By 2019, he developed High-Flyer as a hedge fund concentrated on establishing and utilizing AI trading algorithms. By 2021, High-Flyer specifically used AI in trading. [15] DeepSeek has made its generative artificial open source, indicating its code is easily offered for use, adjustment, and viewing. This consists of permission to access and use the source code, along with style files, for constructing functions. [13]

According to 36Kr, Liang had actually developed a shop of 10,000 Nvidia A100 GPUs, which are utilized to train AI [16], before the United States federal government imposed AI chip restrictions on China. [15]

In April 2023, High-Flyer started an artificial basic intelligence laboratory committed to research study establishing AI tools separate from High-Flyer’s financial organization. [17] [18] In May 2023, with High-Flyer as one of the financiers, the lab became its own business, DeepSeek. [15] [19] [18] Equity capital companies hesitated in supplying financing as it was unlikely that it would have the ability to generate an exit in a short period of time. [15]

After launching DeepSeek-V2 in May 2024, which offered strong performance for a low rate, DeepSeek ended up being known as the driver for China’s AI design price war. It was rapidly dubbed the “Pinduoduo of AI”, and other major tech giants such as ByteDance, Tencent, Baidu, and Alibaba began to cut the price of their AI models to take on the company. Despite the low price charged by DeepSeek, it was rewarding compared to its rivals that were losing money. [20]

DeepSeek is focused on research study and has no detailed prepare for commercialization; [20] this likewise allows its technology to prevent the most rigid arrangements of China’s AI policies, such as needing consumer-facing innovation to abide by the federal government’s controls on information. [3]

DeepSeek’s working with preferences target technical abilities instead of work experience, resulting in many new hires being either current university graduates or designers whose AI professions are less developed. [18] [3] Likewise, the business recruits individuals without any computer technology background to help its innovation comprehend other topics and knowledge locations, including having the ability to generate poetry and carry out well on the infamously hard Chinese college admissions exams (Gaokao). [3]

Development and release history

DeepSeek LLM

On 2 November 2023, DeepSeek released its first series of design, DeepSeek-Coder, which is readily available for free to both scientists and business users. The code for the model was made open-source under the MIT license, with an extra license arrangement (“DeepSeek license”) relating to “open and responsible downstream use” for the design itself. [21]

They are of the very same architecture as DeepSeek LLM detailed listed below. The series consists of 8 models, 4 pretrained (Base) and 4 instruction-finetuned (Instruct). They all have 16K context lengths. The training was as follows: [22] [23] [24]

1. Pretraining: 1.8 T tokens (87% source code, 10% code-related English (GitHub markdown and Stack Exchange), and 3% code-unrelated Chinese).
2. Long-context pretraining: 200B tokens. This extends the context length from 4K to 16K. This produced the Base models.
3. Supervised finetuning (SFT): 2B tokens of instruction information. This produced the Instruct designs.

They were trained on clusters of A100 and H800 Nvidia GPUs, connected by InfiniBand, NVLink, NVSwitch. [22]

On 29 November 2023, DeepSeek released the DeepSeek-LLM series of models, with 7B and 67B specifications in both Base and Chat types (no Instruct was launched). It was established to contend with other LLMs offered at the time. The paper claimed benchmark outcomes greater than many open source LLMs at the time, specifically Llama 2. [26]: section 5 Like DeepSeek Coder, the code for the design was under MIT license, with DeepSeek license for the design itself. [27]

The architecture was essentially the same as those of the Llama series. They used the pre-norm 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-level BPE) and context length of 4096. They trained on 2 trillion tokens of English and Chinese text gotten by deduplicating the Common Crawl. [26]

The Chat variations of the 2 Base designs was also launched concurrently, obtained by training Base by monitored finetuning (SFT) followed by direct policy optimization (DPO). [26]

On 9 January 2024, they launched 2 DeepSeek-MoE models (Base, Chat), each of 16B parameters (2.7 B triggered per token, 4K context length). The training was basically the like DeepSeek-LLM 7B, and was trained on a part of its training dataset. They claimed similar performance with a 16B MoE as a 7B non-MoE. In architecture, it is a variant of the basic sparsely-gated MoE, with “shared specialists” that are always queried, and “routed specialists” that may not be. They discovered this to help with skilled balancing. In standard MoE, some experts can become extremely relied on, while other specialists may be hardly ever used, squandering specifications. Attempting to stabilize the specialists so that they are equally used then causes experts to replicate the same capability. They proposed the shared experts to find out core capabilities that are often used, and let the routed professionals to learn the peripheral capabilities that are hardly ever used. [28]

In April 2024, they released 3 DeepSeek-Math designs specialized for doing math: Base, Instruct, RL. It was trained as follows: [29]

1. Initialize with a formerly pretrained DeepSeek-Coder-Base-v1.5 7B.
2. Further pretrain with 500B tokens (6% DeepSeekMath Corpus, 4% AlgebraicStack, 10% arXiv, 20% GitHub code, 10% Common Crawl). This produced the Base design.
3. Train an instruction-following design by SFT Base with 776K mathematics problems and their tool-use-integrated step-by-step services. This produced the Instruct model.
Reinforcement learning (RL): The reward model was a procedure reward model (PRM) trained from Base according to the Math-Shepherd approach. [30] This reward design was then used to train Instruct utilizing group relative policy optimization (GRPO) on a dataset of 144K math questions “associated to GSM8K and MATH”. The reward model was continuously upgraded during training to prevent reward hacking. This resulted in the RL design.

V2

In May 2024, they released the DeepSeek-V2 series. The series includes 4 designs, 2 base designs (DeepSeek-V2, DeepSeek-V2-Lite) and 2 chatbots (-Chat). The two larger models were trained as follows: [31]

1. Pretrain on a dataset of 8.1 T tokens, where Chinese tokens are 12% more than English ones.
2. Extend context length from 4K to 128K using YaRN. [32] This resulted in DeepSeek-V2.
3. SFT with 1.2 M instances for helpfulness and 0.3 M for safety. This led to DeepSeek-V2-Chat (SFT) which was not released.
4. RL utilizing GRPO in two phases. The very first stage was trained to fix mathematics and coding issues. This stage utilized 1 reward model, trained on compiler feedback (for coding) and ground-truth labels (for math). The 2nd phase was trained to be helpful, safe, and follow guidelines. This phase used 3 benefit models. The helpfulness and security reward models were trained on human preference information. The rule-based benefit design was by hand configured. All trained benefit designs were initialized from DeepSeek-V2-Chat (SFT). This resulted in the launched version of DeepSeek-V2-Chat.

They selected 2-staged RL, since they discovered that RL on reasoning information had “unique attributes” different from RL on general information. For example, RL on reasoning could improve over more training actions. [31]

The 2 V2-Lite models were smaller sized, and qualified likewise, though DeepSeek-V2-Lite-Chat just went through SFT, not RL. They trained the Lite version to assist “additional research study and development on MLA and DeepSeekMoE”. [31]

Architecturally, the V2 designs were significantly customized from the DeepSeek LLM series. They altered the basic attention mechanism by a low-rank approximation called multi-head hidden attention (MLA), and used the mixture of specialists (MoE) variant previously published in January. [28]

The Financial Times reported that it was cheaper than its peers with a rate of 2 RMB for each million output tokens. The University of Waterloo Tiger Lab’s leaderboard ranked DeepSeek-V2 seventh on its LLM ranking. [19]

In June 2024, they released 4 designs in the DeepSeek-Coder-V2 series: V2-Base, V2-Lite-Base, V2-Instruct, V2-Lite-Instruct. They were trained as follows: [35] [note 2]

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

DeepSeek-V2.5 was launched in September and updated in December 2024. It was made by integrating DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct. [36]

V3

In December 2024, they launched a base model DeepSeek-V3-Base and a chat model DeepSeek-V3. The design architecture is basically the very same as V2. They were trained as follows: [37]

1. Pretraining on 14.8 T tokens of a multilingual corpus, mostly English and Chinese. It contained a higher ratio of mathematics and programming than the pretraining dataset of V2.
2. Extend context length two times, from 4K to 32K and then to 128K, using YaRN. [32] This produced DeepSeek-V3-Base.
3. SFT for 2 epochs on 1.5 M samples of reasoning (mathematics, shows, logic) and non-reasoning (creative writing, roleplay, easy question answering) information. Reasoning information was generated by “professional models”. Non-reasoning data was generated by DeepSeek-V2.5 and checked by humans. – The “skilled designs” were trained by beginning with an unspecified base design, then SFT on both information, and synthetic information generated by an internal DeepSeek-R1 model. The system timely asked the R1 to show and verify during thinking. Then the expert models were RL utilizing an unspecified reward function.
– Each professional design was trained to generate simply artificial thinking data in one particular domain (mathematics, shows, logic).
– Expert designs were utilized, instead of R1 itself, considering that the output from R1 itself suffered “overthinking, poor format, and extreme length”.

4. Model-based benefit models were made by beginning with a SFT checkpoint of V3, then finetuning on human choice information containing both last reward and chain-of-thought leading to the last reward. The reward model produced benefit signals for both questions with objective however free-form responses, and questions without unbiased answers (such as imaginative writing).
5. A SFT checkpoint of V3 was trained by GRPO utilizing both reward models and rule-based benefit. The rule-based benefit was computed for mathematics problems with a final answer (put in a box), and for programming issues by unit tests. This produced DeepSeek-V3.

The DeepSeek group performed comprehensive low-level engineering to accomplish efficiency. They utilized mixed-precision math. Much of the forward pass was performed in 8-bit drifting point numbers (5E2M: 5-bit exponent and 2-bit mantissa) rather than the basic 32-bit, needing unique GEMM regimens to accumulate precisely. They used a custom-made 12-bit float (E5M6) for only the inputs to the linear layers after the attention modules. Optimizer states remained in 16-bit (BF16). They lessened the communication latency by overlapping thoroughly computation and interaction, such as devoting 20 streaming multiprocessors out of 132 per H800 for only inter-GPU interaction. They reduced communication by rearranging (every 10 minutes) the exact maker each specialist was on in order to avoid certain machines being queried more often than the others, including auxiliary load-balancing losses to the training loss function, and other load-balancing methods. [37]

After training, it was deployed on H800 clusters. The H800 cards within a cluster are connected by NVLink, and the clusters are linked by InfiniBand. [37]

Benchmark tests show that DeepSeek-V3 exceeded Llama 3.1 and Qwen 2.5 whilst matching GPT-4o and Claude 3.5 Sonnet. [18] [39] [40] [41]

R1

On 20 November 2024, DeepSeek-R1-Lite-Preview became available through DeepSeek’s API, along with through a chat interface after logging in. [42] [43] [note 3] It was trained for rational reasoning, mathematical reasoning, and real-time problem-solving. DeepSeek claimed that it exceeded performance of OpenAI o1 on criteria such as American Invitational Mathematics Examination (AIME) and MATH. [44] However, The Wall Street Journal mentioned when it utilized 15 problems from the 2024 edition of AIME, the o1 model reached an option much faster than DeepSeek-R1-Lite-Preview. [45]

On 20 January 2025, DeepSeek released DeepSeek-R1 and DeepSeek-R1-Zero. [46] Both were initialized from DeepSeek-V3-Base, and share its architecture. The business also launched some “DeepSeek-R1-Distill” models, which are not initialized on V3-Base, however instead are initialized from other pretrained open-weight designs, consisting of LLaMA and Qwen, then fine-tuned on synthetic data generated by R1. [47]

A conversation between User and Assistant. The user asks a concern, and the Assistant resolves it. The assistant first considers the reasoning process in the mind and after that provides the user with the answer. The reasoning process and response are enclosed within and tags, respectively, i.e., thinking procedure here answer here. User:. Assistant:

DeepSeek-R1-Zero was trained solely using GRPO RL without SFT. Unlike previous variations, they used no model-based benefit. All reward functions were rule-based, “generally” of two types (other types were not defined): precision rewards and format benefits. Accuracy reward was inspecting whether a boxed response is appropriate (for math) or whether a code passes tests (for shows). Format reward was inspecting whether the design puts its thinking trace within … [47]

As R1-Zero has problems with readability and blending languages, R1 was trained to resolve these problems and more improve thinking: [47]

1. SFT DeepSeek-V3-Base on “thousands” of “cold-start” information all with the basic format of|special_token|| special_token|summary >.
2. Apply the same RL process as R1-Zero, however likewise with a “language consistency reward” to encourage it to react monolingually. This produced an internal design not released.
3. Synthesize 600K thinking information from the internal design, 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.
4. SFT DeepSeek-V3-Base on the 800K synthetic data for 2 dates.
5. GRPO RL with rule-based benefit (for reasoning jobs) and model-based benefit (for non-reasoning jobs, helpfulness, and harmlessness). This produced DeepSeek-R1.

Distilled designs were trained by SFT on 800K information synthesized from DeepSeek-R1, in a comparable way as action 3 above. They were not trained with RL. [47]

Assessment and responses

DeepSeek released its AI Assistant, which uses the V3 design as a chatbot app for Apple IOS and Android. By 27 January 2025 the app had actually exceeded ChatGPT as the highest-rated totally free app on the iOS App Store in the United States; its chatbot reportedly responds to questions, resolves logic issues and writes computer system programs on par with other chatbots on the market, according to benchmark tests used by American AI business. [3]

DeepSeek-V3 uses considerably less resources compared to its peers; for instance, whereas the world’s leading AI business train their chatbots with supercomputers using as lots of as 16,000 graphics processing systems (GPUs), if not more, DeepSeek claims to require only about 2,000 GPUs, specifically the H800 series chip from Nvidia. [37] It was trained in around 55 days at an expense of US$ 5.58 million, [37] which is roughly one tenth of what United States tech huge Meta invested developing its most current AI innovation. [3]

DeepSeek’s competitive efficiency at fairly minimal expense has been recognized as possibly challenging the global supremacy of American AI designs. [48] Various publications and news media, such as The Hill and The Guardian, described the release of its chatbot as a “Sputnik moment” for American AI. [49] [50] The performance of its R1 model was apparently “on par with” one of OpenAI’s latest models when used for jobs such as mathematics, coding, and natural language reasoning; [51] echoing other commentators, American Silicon Valley venture capitalist Marc Andreessen similarly described R1 as “AI’s Sputnik moment”. [51]

DeepSeek’s founder, Liang Wenfeng has actually been compared to Open AI CEO Sam Altman, with CNN calling him the Sam Altman of China and an evangelist for AI. [52] Chinese state media commonly applauded DeepSeek as a nationwide possession. [53] [54] On 20 January 2025, China’s Premier Li Qiang welcomed Liang Wenfeng to his symposium with experts and asked him to provide opinions and suggestions on a draft for comments of the annual 2024 federal government work report. [55]

DeepSeek’s optimization of minimal resources has actually highlighted potential limits of United States sanctions on China’s AI development, which consist of export constraints on sophisticated AI chips to China [18] [56] The success of the business’s AI models subsequently “triggered market chaos” [57] and triggered shares in major international technology business to plunge on 27 January 2025: Nvidia’s stock fell by as much as 17-18%, [58] as did the stock of rival Broadcom. Other tech firms likewise sank, consisting of Microsoft (down 2.5%), Google’s owner Alphabet (down over 4%), and Dutch chip devices maker ASML (down over 7%). [51] An international selloff of technology stocks on Nasdaq, prompted by the release of the R1 model, had actually resulted in record losses of about $593 billion in the market capitalizations of AI and hardware business; [59] by 28 January 2025, a total of $1 trillion of value was rubbed out American stocks. [50]

Leading figures in the American AI sector had combined responses to DeepSeek’s success and performance. [60] Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman-whose business are associated with the United States government-backed “Stargate Project” to develop American AI infrastructure-both called DeepSeek “extremely impressive”. [61] [62] American President Donald Trump, who announced The Stargate Project, called DeepSeek a wake-up call [63] and a positive development. [64] [50] [51] [65] Other leaders in the field, consisting of Scale AI CEO Alexandr Wang, Anthropic cofounder and CEO Dario Amodei, and Elon Musk revealed hesitation of the app’s efficiency or of the sustainability of its success. [60] [66] [67] Various companies, consisting of Amazon Web Services, Toyota, and Stripe, are seeking to utilize the model in their program. [68]

On 27 January 2025, DeepSeek restricted its new user registration to contact number from mainland China, email addresses, or Google account logins, following a “massive” cyberattack interfered with the correct performance of its servers. [69] [70]

Some sources have actually observed that the official application shows user interface (API) variation of R1, which ranges from servers located in China, uses censorship systems for topics that are considered politically sensitive for the government of China. For example, the design refuses to answer concerns about the 1989 Tiananmen Square demonstrations and massacre, persecution of Uyghurs, comparisons between Xi Jinping and Winnie the Pooh, or human rights in China. [71] [72] [73] The AI might at first produce a response, however then deletes it soon later on and changes it with a message such as: “Sorry, that’s beyond my present scope. Let’s speak about something else.” [72] The integrated censorship mechanisms and constraints can just be gotten rid of to a minimal degree in the open-source version of the R1 model. If the “core socialist worths” specified by the Chinese Internet regulative authorities are discussed, or the political status of Taiwan is raised, discussions are terminated. [74] When evaluated by NBC News, DeepSeek’s R1 explained Taiwan as “an inalienable part of China’s territory,” and specified: “We securely oppose any kind of ‘Taiwan self-reliance’ separatist activities and are committed to accomplishing the total reunification of the motherland through tranquil methods.” [75] In January 2025, Western researchers were able to deceive DeepSeek into giving certain responses to a few of these subjects by asking for in its answer to swap certain letters for similar-looking numbers. [73]

Security and privacy

Some experts fear that the federal government of China could use the AI system for foreign influence operations, spreading disinformation, surveillance and the development of cyberweapons. [76] [77] [78] DeepSeek’s personal privacy terms say “We keep the information we collect in safe servers located in individuals’s Republic of China … We may collect your text or audio input, timely, uploaded files, feedback, chat history, or other content that you offer to our design and Services”. Although the data storage and collection policy follows ChatGPT’s privacy policy, [79] a Wired post reports this as security issues. [80] In action, the Italian data protection authority is looking for additional details on DeepSeek’s collection and usage of individual information, and the United States National Security Council revealed that it had started a nationwide security evaluation. [81] [82] Taiwan’s federal government banned the use of DeepSeek at federal government ministries on security grounds and South Korea’s Personal Information Protection Commission opened a query into DeepSeek’s usage of individual info. [83]

Artificial intelligence market in China.

Notes

^ a b c The variety of heads does not equivalent the variety of KV heads, due to GQA.
^ Inexplicably, the model named DeepSeek-Coder-V2 Chat in the paper was launched as DeepSeek-Coder-V2-Instruct in HuggingFace.
^ At that time, the R1-Lite-Preview needed choosing “Deep Think allowed”, and every user could use it only 50 times a day.
References

^ Gibney, Elizabeth (23 January 2025). “China’s cheap, open AI design DeepSeek thrills scientists”. Nature. doi:10.1038/ d41586-025-00229-6. ISSN 1476-4687. PMID 39849139.
^ a b Vincent, James (28 January 2025). “The DeepSeek panic exposes an AI world all set to blow”. The Guardian.
^ a b c d e f g Metz, Cade; Tobin, Meaghan (23 January 2025). “How Chinese A.I. Start-Up DeepSeek Is Taking On Silicon Valley Giants”. The New York Times. ISSN 0362-4331. Retrieved 27 January 2025.
^ Cosgrove, Emma (27 January 2025). “DeepSeek’s more affordable designs and weaker chips call into concern trillions in AI infrastructure costs”. Business Insider.
^ Mallick, Subhrojit (16 January 2024). “Biden admin’s cap on GPU exports may strike India’s AI ambitions”. The Economic Times. Retrieved 29 January 2025.
^ Saran, Cliff (10 December 2024). “Nvidia examination signals expanding of US and China chip war|Computer Weekly”. Computer Weekly. Retrieved 27 January 2025.
^ Sherman, Natalie (9 December 2024). “Nvidia targeted by China in brand-new chip war probe”. BBC. Retrieved 27 January 2025.
^ a b c Metz, Cade (27 January 2025). “What is DeepSeek? And How Is It Upending A.I.?”. The New York Times. ISSN 0362-4331. Retrieved 27 January 2025.
^ Field, Hayden (27 January 2025). “China’s DeepSeek AI dethrones ChatGPT on App Store: Here’s what you need to know”. CNBC.
^ Picchi, Aimee (27 January 2025). “What is DeepSeek, and why is it triggering Nvidia and other stocks to plunge?”. CBS News.
^ Zahn, Max (27 January 2025). “Nvidia, Microsoft shares tumble as China-based AI app DeepSeek hammers tech giants”. ABC News. Retrieved 27 January 2025.
^ Roose, Kevin (28 January 2025). “Why DeepSeek Could Change What Silicon Valley Believe About A.I.” The New York Times. ISSN 0362-4331. Retrieved 28 January 2025.
^ a b Romero, Luis E. (28 January 2025). “ChatGPT, DeepSeek, Or Llama? Meta’s LeCun Says Open-Source Is The Key”. Forbes.
^ Chen, Caiwei (24 January 2025). “How a top Chinese AI design overcame US sanctions”. MIT Technology Review. Archived from the original on 25 January 2025. Retrieved 25 January 2025.
^ a b c d Ottinger, Lily (9 December 2024). “Deepseek: From Hedge Fund to Frontier Model Maker”. ChinaTalk. Archived from the original on 28 December 2024. Retrieved 28 December 2024.
^ Leswing, Kif (23 February 2023). “Meet the $10,000 Nvidia chip powering the race for A.I.” CNBC. Retrieved 30 January 2025.
^ Yu, Xu (17 April 2023).” [Exclusive] Chinese Quant Hedge Fund High-Flyer Won’t Use AGI to Trade Stocks, MD Says”. Yicai Global. Archived from the original on 31 December 2023. Retrieved 28 December 2024.
^ a b c d e Jiang, Ben; Perezi, Bien (1 January 2025). “Meet DeepSeek: the Chinese start-up that is changing how AI models are trained”. South China Morning Post. Archived from the initial on 22 January 2025. Retrieved 1 January 2025.
^ a b McMorrow, Ryan; Olcott, Eleanor (9 June 2024). “The Chinese quant fund-turned-AI leader”. Financial Times. Archived from the original on 17 July 2024. Retrieved 28 December 2024.
^ a b Schneider, Jordan (27 November 2024). “Deepseek: The Quiet Giant Leading China’s AI Race”. ChinaTalk. Retrieved 28 December 2024.
^ “DeepSeek-Coder/LICENSE-MODEL at main · deepseek-ai/DeepSeek-Coder”. GitHub. Archived from the initial on 22 January 2025. Retrieved 24 January 2025.
^ a b c Guo, Daya; Zhu, Qihao; Yang, Dejian; Xie, Zhenda; Dong, Kai; Zhang, Wentao; Chen, Guanting; Bi, Xiao; Wu, Y. (26 January 2024), DeepSeek-Coder: When the Large Language Model Meets Programming – The Rise of Code Intelligence, arXiv:2401.14196.
^ “DeepSeek Coder”. deepseekcoder.github.io. Retrieved 27 January 2025.
^ deepseek-ai/DeepSeek-Coder, DeepSeek, 27 January 2025, retrieved 27 January 2025.
^ “deepseek-ai/deepseek-coder -5.7 bmqa-base · Hugging Face”. huggingface.co. Retrieved 27 January 2025.
^ a b c d DeepSeek-AI; Bi, Xiao; Chen, Deli; Chen, Guanting; Chen, Shanhuang; Dai, Damai; Deng, Chengqi; Ding, Honghui; Dong, Kai (5 January 2024), DeepSeek LLM: Scaling Open-Source Language Models with Longtermism, arXiv:2401.02954.
^ deepseek-ai/DeepSeek-LLM, DeepSeek, 27 January 2025, obtained 27 January 2025.
^ a b Dai, Damai; Deng, Chengqi; Zhao, Chenggang; Xu, R. X.; Gao, Huazuo; Chen, Deli; Li, Jiashi; Zeng, Wangding; Yu, Xingkai (11 January 2024), DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models, arXiv:2401.06066.
^ Shao, Zhihong; Wang, Peiyi; Zhu, Qihao; Xu, Runxin; Song, Junxiao; Bi, Xiao; Zhang, Haowei; Zhang, Mingchuan; Li, Y. K. (27 April 2024), DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models, arXiv:2402.03300.
^ Wang, Peiyi; Li, Lei; Shao, Zhihong; Xu, R. X.; Dai, Damai; Li, Yifei; Chen, Deli; Wu, Y.; Sui, Zhifang (19 February 2024), Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations, arXiv:2312.08935. ^ a b c d DeepSeek-AI; Liu, Aixin; Feng, Bei; Wang, Bin; Wang, Bingxuan; Liu, Bo; Zhao, Chenggang; Dengr, Chengqi; Ruan, Chong (19 June 2024), DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model, arXiv:2405.04434.
^ a b Peng, Bowen; Quesnelle, Jeffrey; Fan, Honglu; Shippole, Enrico (1 November 2023), YaRN: Efficient Context Window Extension of Large Language Models, arXiv:2309.00071.
^ “config.json · deepseek-ai/DeepSeek-V 2-Lite at primary”. huggingface.co. 15 May 2024. Retrieved 28 January 2025.
^ “config.json · deepseek-ai/DeepSeek-V 2 at primary”. huggingface.co. 6 May 2024. Retrieved 28 January 2025.
^ DeepSeek-AI; Zhu, Qihao; Guo, Daya; Shao, Zhihong; Yang, Dejian; Wang, Peiyi; Xu, Runxin; Wu, Y.; Li, Yukun (17 June 2024), DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence, arXiv:2406.11931.
^ “deepseek-ai/DeepSeek-V 2.5 · Hugging Face”. huggingface.co. 3 January 2025. Retrieved 28 January 2025.
^ a b c d e f g DeepSeek-AI; Liu, Aixin; Feng, Bei; Xue, Bing; Wang, Bingxuan; Wu, Bochao; Lu, Chengda; Zhao, Chenggang; Deng, Chengqi (27 December 2024), DeepSeek-V3 Technical Report, arXiv:2412.19437.
^ “config.json · deepseek-ai/DeepSeek-V 3 at main”. huggingface.co. 26 December 2024. Retrieved 28 January 2025.
^ Jiang, Ben (27 December 2024). “Chinese start-up DeepSeek’s new AI model surpasses Meta, OpenAI products”. South China Morning Post. Archived from the initial on 27 December 2024. Retrieved 28 December 2024.
^ Sharma, Shubham (26 December 2024). “DeepSeek-V3, ultra-large open-source AI, surpasses Llama and Qwen on launch”. VentureBeat. Archived from the original on 27 December 2024. Retrieved 28 December 2024.
^ Wiggers, Kyle (26 December 2024). “DeepSeek’s new AI model appears to be one of the very best ‘open’ oppositions yet”. TechCrunch. Archived from the initial on 2 January 2025. Retrieved 31 December 2024.
^ “Deepseek Log in page”. DeepSeek. Retrieved 30 January 2025.
^ “News|DeepSeek-R1-Lite Release 2024/11/20: DeepSeek-R1-Lite-Preview is now live: unleashing supercharged thinking power!”. DeepSeek API Docs. Archived from the initial on 20 November 2024. Retrieved 28 January 2025.
^ Franzen, Carl (20 November 2024). “DeepSeek’s very first reasoning model R1-Lite-Preview turns heads, beating OpenAI o1 performance”. VentureBeat. Archived from the original on 22 November 2024. Retrieved 28 December 2024.
^ Huang, Raffaele (24 December 2024). “Don’t Look Now, but China’s AI Is Catching Up Fast”. The Wall Street Journal. Archived from the original on 27 December 2024. Retrieved 28 December 2024.
^ “Release DeepSeek-R1 · deepseek-ai/DeepSeek-R1@23807ce”. GitHub. Archived from the original on 21 January 2025. Retrieved 21 January 2025.
^ a b c d DeepSeek-AI; Guo, Daya; Yang, Dejian; Zhang, Haowei; Song, Junxiao; Zhang, Ruoyu; Xu, Runxin; Zhu, Qihao; Ma, Shirong (22 January 2025), DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, arXiv:2501.12948.
^ “Chinese AI startup DeepSeek surpasses ChatGPT on Apple App Store”. Reuters. 27 January 2025. Retrieved 27 January 2025.
^ Wade, David (6 December 2024). “American AI has actually reached its Sputnik moment”. The Hill. Archived from the original on 8 December 2024. Retrieved 25 January 2025.
^ a b c Milmo, Dan; Hawkins, Amy; Booth, Robert; Kollewe, Julia (28 January 2025). “‘ Sputnik minute’: $1tn rubbed out US stocks after Chinese company reveals AI chatbot” – via The Guardian.
^ a b c d Hoskins, Peter; Rahman-Jones, Imran (27 January 2025). “Nvidia shares sink as Chinese AI app spooks markets”. BBC. Retrieved 28 January 2025.
^ Goldman, David (27 January 2025). “What is DeepSeek, the Chinese AI start-up that shook the tech world?|CNN Business”. CNN. Retrieved 29 January 2025.
^ “DeepSeek postures a difficulty to Beijing as much as to Silicon Valley”. The Economist. 29 January 2025. ISSN 0013-0613. Retrieved 31 January 2025.
^ Paul, Katie; Nellis, Stephen (30 January 2025). “Chinese state-linked accounts hyped DeepSeek AI launch ahead of US stock rout, Graphika says”. Reuters. Retrieved 30 January 2025.
^ 澎湃新闻 (22 January 2025). “量化巨头幻方创始人梁文锋参加总理座谈会并发言 ， 他还创办了” AI界拼多多””. finance.sina.com.cn. Retrieved 31 January 2025.
^ Shilov, Anton (27 December 2024). “Chinese AI company’s AI design advancement highlights limits of US sanctions”. Tom’s Hardware. Archived from the initial on 28 December 2024. Retrieved 28 December 2024.
^ “DeepSeek updates – Chinese AI chatbot stimulates US market chaos, cleaning $500bn off Nvidia”. BBC News. Retrieved 27 January 2025.
^ Nazareth, Rita (26 January 2025). “Stock Rout Gets Ugly as Nvidia Extends Loss to 17%: Markets Wrap”. Bloomberg. Retrieved 27 January 2025.
^ Carew, Sinéad; Cooper, Amanda; Banerjee, Ankur (27 January 2025). “DeepSeek sparks global AI selloff, Nvidia losses about $593 billion of value”. Reuters.
^ a b Sherry, Ben (28 January 2025). “DeepSeek, Calling It ‘Impressive’ but Staying Skeptical”. Inc. Retrieved 29 January 2025.
^ Okemwa, Kevin (28 January 2025). “Microsoft CEO Satya Nadella touts DeepSeek’s open-source AI as “very impressive”: “We need to take the advancements out of China very, extremely seriously””. Windows Central. Retrieved 28 January 2025.
^ Nazzaro, Miranda (28 January 2025). “OpenAI’s Sam Altman calls DeepSeek model ‘outstanding'”. The Hill. Retrieved 28 January 2025.
^ Dou, Eva; Gregg, Aaron; Zakrzewski, Cat; Tiku, Nitasha; Najmabadi, Shannon (28 January 2025). “Trump calls China’s DeepSeek AI app a ‘wake-up call’ after tech stocks slide”. The Washington Post. Retrieved 28 January 2025.
^ Habeshian, Sareen (28 January 2025). “Johnson slams China on AI, Trump calls DeepSeek development “favorable””. Axios.
^ Karaian, Jason; Rennison, Joe (27 January 2025). “China’s A.I. Advances Spook Big Tech Investors on Wall Street” – through NYTimes.com.
^ Sharma, Manoj (6 January 2025). “Musk dismisses, Altman applauds: What leaders say on DeepSeek’s interruption”. Fortune India. Retrieved 28 January 2025.
^ “Elon Musk ‘concerns’ DeepSeek’s claims, recommends enormous Nvidia GPU infrastructure”. Financialexpress. 28 January 2025. Retrieved 28 January 2025.
^ Kim, Eugene. “Big AWS customers, including Stripe and Toyota, are pestering the cloud giant for access to DeepSeek AI designs”. Business Insider.
^ Kerr, Dara (27 January 2025). “DeepSeek hit with ‘massive’ cyber-attack after AI chatbot tops app shops”. The Guardian. Retrieved 28 January 2025.
^ Tweedie, Steven; Altchek, Ana. “DeepSeek temporarily limited brand-new sign-ups, pointing out ‘large-scale destructive attacks'”. Business Insider.
^ Field, Matthew; Titcomb, James (27 January 2025). “Chinese AI has triggered a $1 trillion panic – and it doesn’t care about totally free speech”. The Daily Telegraph. ISSN 0307-1235. Retrieved 27 January 2025.
^ a b Steinschaden, Jakob (27 January 2025). “DeepSeek: This is what live censorship appears like in the Chinese AI chatbot”. Trending Topics. Retrieved 27 January 2025.
^ a b Lu, Donna (28 January 2025). “We tried DeepSeek. It worked well, until we asked it about Tiananmen Square and Taiwan”. The Guardian. ISSN 0261-3077. Retrieved 30 January 2025.
^ “The Guardian view on an international AI race: geopolitics, innovation and the increase of chaos”. The Guardian. 26 January 2025. ISSN 0261-3077. Retrieved 27 January 2025.
^ Yang, Angela; Cui, Jasmine (27 January 2025). “Chinese AI DeepSeek shocks Silicon Valley, offering the AI race its ‘Sputnik minute'”. NBC News. Retrieved 27 January 2025.
^ Kimery, Anthony (26 January 2025). “China’s DeepSeek AI postures formidable cyber, data personal privacy threats”. Biometric Update. Retrieved 27 January 2025.
^ Booth, Robert; Milmo, Dan (28 January 2025). “Experts prompt care over usage of Chinese AI DeepSeek”. The Guardian. ISSN 0261-3077. Retrieved 28 January 2025.
^ Hornby, Rael (28 January 2025). “DeepSeek’s success has painted a substantial TikTok-shaped target on its back”. LaptopMag. Retrieved 28 January 2025.
^ “Privacy policy”. Open AI. Retrieved 28 January 2025.
^ Burgess, Matt; Newman, Lily Hay (27 January 2025). “DeepSeek’s Popular AI App Is Explicitly Sending US Data to China”. Wired. ISSN 1059-1028. Retrieved 28 January 2025.
^ “Italy regulator seeks details from DeepSeek on information security”. Reuters. 28 January 2025. Retrieved 28 January 2025.
^ Shalal, Andrea; Shepardson, David (28 January 2025). “White House assesses impact of China AI app DeepSeek on national security, authorities says”. Reuters. Retrieved 28 January 2025.