DeepSeek

以极致性价比的open-source large model矩阵席卷全球

large language model

https://www.deepseek.com/

Visit Website

In-depth Report

DeepSeek is one of the most influential companies in China's AI field. It was incubated by Magic Square Quantitative in 2023. With its extremely cost-effective API pricing and open source strategy, it occupies a unique position in the global large model competition. Its core product matrix includes general conversation model V3.2, deep reasoning model R1, professional code model Coder V2, and the soon-to-be-released trillion-parameter multi-modal model V4. In April 2026, DeepSeek was revealed to have launched its first round of financing since its establishment, with a valuation of more than US$10 billion, marking the official entry of this technology-driven company into the capital market.
The parent company of DeepSeek is Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., which was incubated and established in 2023 by the quantitative private equity giant Magic Square Quantitative. Founder Liang Wenfeng is also the founder of Huanfang Quantitative, owning 84.29% of the company's shares and nearly 100% of the voting rights. Huanfang Quantitative is the leading quantitative private equity firm in China, with a management scale of over RMB 70 billion, an average return rate of 56.6% in 2025, and an estimated annual revenue of approximately RMB 5 billion according to industry practice. This provides sufficient financial support for DeepSeek’s entire R&D investment in the first three years. In the first three years since its establishment, DeepSeek has never received any external equity financing. All computing power procurement, team salaries and R&D expenditures were borne by Huifang Quantification, with a cumulative investment of approximately hundreds of millions of dollars in three years. This "self-generation" model allows DeepSeek to maintain a high degree of technical independence, but it also faces pressure from brain drain - V2 architecture contributor Luo Fuli joined Xiaomi, and GRPO algorithm core author Guo Daya joined ByteDance, prompting the company to launch its first round of financing in April 2026. According to media reports such as The Information, DeepSeek plans to raise at least US$300 million at a valuation of no less than US$10 billion and release about 3% of its equity. Its core purpose is to establish a market-based pricing anchor for employee options. Compared with domestic peers, DeepSeek’s valuation is relatively low—Zhipu AI has a market value of US$50.7 billion, MiniMax has a market value of US$34.4 billion, and Dark Side of the Moon is valued at US$18 billion. DeepSeek's brand slogan is "Exploring the Unexplored", and the open source warehouse @deepseek-ai on GitHub has extremely high community influence. All its main models are open source using the MIT license or the Apache 2.0 license, allowing commercial use and secondary development.
DeepSeek's current product matrix covers four major directions: general dialogue, deep reasoning, code generation and multi-modal understanding. In terms of general dialogue, DeepSeek-V3.2 is the latest flagship model with a total parameter of 671 billion, an activation parameter of about 37B, and supports a 128K context window (about 200,000 words). The model adopts MoE (Mixed Experts) sparse architecture, and each token only activates some parameters, achieving "the capabilities of a large model and the cost of a small model." The V3.2 version adds new Agent capabilities, strengthens the agent function and integrates thinking and reasoning, and adopts an intelligent tool calling mechanism of "reason first and then call". In addition, V3.2 is also the first open source model to implement "thinking integration into tool invocation". Together with DSA sparse attention technology, it reduces long text calculation costs by 50% to 70%. In terms of deep reasoning, DeepSeek-R1 is the core reasoning model, which adopts a reinforcement learning-driven reasoning paradigm and supports visual thinking chains and self-correction capabilities. The accuracy rate on the mathematics competition AIME benchmark reached 87.5%, and the code benchmark LiveCodeBench score was 83.3. The accuracy rate on complex tasks is about 30% higher than that of ordinary models. R1 also supports model distillation, which can transfer reasoning capabilities to small models such as Qwen and Llama. In terms of code generation, DeepSeek Coder V2 supports more than 20 programming languages, supports IDE integration, and has a higher success rate in bug locating than GPT-4o. Chinese language capability is DeepSeek’s outstanding advantage. Special optimization has been carried out in ancient Chinese translation, poetry creation, technical documentation, etc., supporting more than 100 languages, and the quality of Chinese-English translation is leading among domestic models. The upcoming DeepSeek V4 is the product that is currently receiving the most attention. According to the leaked specification information, the total parameters of V4 will reach approximately 1 trillion, the activation parameters will remain unchanged at approximately 37B, the context window will be expanded to 1M+ tokens, and it will support native multi-modality (text, pictures, video, audio). Core technologies include the Engram conditional memory module (O(1) constant time retrieval, long context accuracy increased by 12.8%) and mHC manifold constraint technology (solve the stability problem of trillions of parameter training). The training chip of V4 has switched from NVIDIA to domestic chips such as Huawei Ascend and Cambrian. It is expected that the coding capability SWE-bench will reach more than 80%, which is the same as Claude Opus 4.6 and GPT-5.4.
DeepSeek's pricing strategy is known for its "ultimate cost performance" and is its core competitive weapon. The API price of the V3.2 model (deepseek-chat) is $0.27/million tokens for input and $1.10/million tokens for output. The R1 model (deepseek-reasoner) inputs $0.55/million tokens and outputs $2.19/million tokens. In RMB, the V3.2 input is as low as 0.2 to 2 yuan per million tokens, and the R1 input is 1 to 4 yuan per million tokens. In addition, DeepSeek provides peak-shifting discounts, with prices as low as 50% between 00:30 and 08:30 Beijing time. Compared with competing products, GPT-5 is priced at $1.25 for input and $10.00 for output, which is 4.6 times and 9.1 times that of DeepSeek V3.2. Claude Sonnet 4.5 is input $3.00 and output $15.00, which are 11 times and 13.6 times that of V3.2. Even the cheapest Llama 4 Maverick hosting plan is only about 30% cheaper than V3.2. Taking a moderately used startup team as an example (average daily input of 1 million + 500,000 output tokens), the monthly fee for using V3.2 is about $24.60, while GPT-5 is $187.50, and Claude Sonnet 4.5 is as high as $315.00. DeepSeek adopts a free value-added model: the web version and App provide free conversation services, and the API is pay-as-you-go. Enterprise customers can integrate through cloud services such as Huawei Cloud, Alibaba Cloud, AWS Bedrock, etc., or choose private deployment (need to provide their own hardware). This multi-level business model takes into account both C-side user growth and B-side monetization.
Positive reviews focus on several areas. Developers generally believe that DeepSeek's price/performance ratio is unmatched, and its API price is only 15% to 50% of OpenAI's, making it suitable for production environments with high concurrency and large-scale calls. Chinese processing capabilities are outstanding among domestic models, and the quality of ancient Chinese translation and Chinese writing even exceeds that of some international models. The open source strategy has been widely praised by the technology community. The MIT protocol allows commercial use and secondary development, lowering the technical threshold for many teams. In programming scenarios, DeepSeek's code generation and debugging capabilities are recognized by developers, especially in mainstream languages such as Python and Java. Negative feedback cannot be ignored either. Some users reported that the server was unstable during peak periods, with occasional response delays or timeouts. The timeliness of the knowledge base is a shortcoming. The training data will be available until around May 2025. It needs to cooperate with online searches to obtain the latest information. The multi-modal capability is currently weak and lacks image and audio input functions (V4 is expected to complete this). In terms of English deep reasoning, the R1 model is still slightly inferior to OpenAI o1/o3 and Claude 4. Although the hallucination problem has been significantly improved compared to earlier versions, occasional cases still exist.
DeepSeek detonated the global AI circle with its R1 inference model in early 2025. After the release of R1 in January 2025, NVIDIA's stock price fell sharply, and the global market re-examined the technical strength of Chinese AI companies. International media evaluated DeepSeek as the "OpenAI challenger to Chinese AI" and believed that it achieved performance close to top models with extremely low training costs, breaking the industry narrative that "computing power is everything". Chinese technology media’s evaluation is more in-depth. 36Kr pointed out in an analysis on April 20, 2026 that DeepSeek’s valuation of US$10 billion in this round is “low”, which is essentially an entry barrier set for employee option pricing. Qubit reported that Liang Wenfeng personally invested approximately 20 billion yuan in the first round of financing, which was both a reassurance to external investors and a declaration of personal belief. At the technical level, DeepSeek's MoE architecture and open source strategy are widely regarded as important factors in promoting industry progress. Its GRPO (Group Relative Policy Optimization) algorithm has been used for reference by many competitors, and its model distillation solution has promoted the overall improvement of small model reasoning capabilities. DeepSeek is considered to be the most representative technical force in the "involution" wave of domestic large models in 2025-2026.
Content moderation is the most controversial issue. As a company registered in China, DeepSeek's model has a censorship mechanism on politically sensitive topics, which has triggered a lot of discussion in overseas communities such as Reddit and Hacker News. Some users believe that this limits the usefulness of the model, especially since it may face trust barriers in international marketing. Dependence on computing power is a structural risk. Although V4 switches to domestic chips, the supply of high-end AI chips is still restricted by geopolitical factors. The U.S. export restrictions on AI chips to China pose a potential threat to DeepSeek’s long-term computing power expansion. Brain drain is an operational challenge. The departure of core architecture contributors and algorithm authors has exposed the fierce competition in the AI talent market. An important purpose of the first round of financing is to retain talents through option incentives. Commercial sustainability also deserves attention. Although the strategy of seizing the market at extremely low prices has effectively expanded the user base, its long-term profitability is questionable. After shifting from quantitative financial support to independent operations, DeepSeek needs to prove its commercial capabilities.
DeepSeek is suitable for several clear usage scenarios: cost-sensitive developers and start-up teams who need large-scale API calls but want to control expenses; Chinese content creators and office users, especially scenarios involving ancient Chinese translation, technical documents, and official document writing; academic researchers and those who require mathematical reasoning, the R1 model performs outstandingly in mathematical and logical reasoning; enterprise customers who require privatized deployment, the MIT open source protocol provides sufficient freedom of customization. Scenarios that are not suitable include: users with strong demand for multi-modality (image generation, voice interaction); international business scenarios that need to deal with politically sensitive topics; professional fields with extremely high requirements for in-depth reasoning in English; users who pursue the timeliness of cutting-edge knowledge. In terms of alternatives, the domestic GLM series, Tongyi Qianwen, Wenxin Yiyan, etc. can be considered; the main competitors in the international market are the OpenAI GPT series, Anthropic Claude series and Google Gemini series. For users who pursue the ultimate cost-effectiveness, Meta’s Llama series of open source models are also optional.
DeepSeek has become the most influential Chinese AI company in the world with its strategy of "breaking down barriers with open source and leveraging the market with low prices." The soon-to-be-released trillion-parameter V4 model will further raise the competitive threshold, and the first round of financing with a valuation of US$10 billion marks the company's transition from a technical experiment to a regular commercial army. Its biggest advantage has always been cost-effectiveness and open source ecology, while its biggest challenge is how to achieve sustainable business growth while maintaining technological independence.

User Reviews

PYoung_2020

—

刚看到DeepSeek要融资的消息，100亿估值真的太低了，跟智谱507亿比差了好几倍。不过梁文锋本来就不缺钱，这轮融资大概率是为了给员工发期权留住人。毕竟核心骨干已经走了好几个了，再不激励真的要被大厂挖光了。

OliviaTorresZ

—

V4到底什么时候发啊，等了快半年了，每次说快了快了结果又鸽。

Edward.Price_66

—

作为一个重度DeepSeek API用户，V3.2的性价比真的没话说。我们公司每个月调用大概2000万token，换成GPT-5要花大几千美金，用DeepSeek才不到200块。代码生成质量虽然比Claude Opus差一点，但日常开发完全够用了，省下来的钱多招两个人不香吗。

ChristopherJones

—

DeepSeek R1的数学推理确实强，AIME竞赛题能拿87.5%准确率这个数据不是吹的，我自己测试了好几道题确实比GPT-4o算得准。不过英文写作质量还是差点意思，输出格式有时候乱七八糟的。

CryptoFi2024

—

V4要是真的能做到万亿参数还保持这个价格，那OpenAI和Anthropic真的要慌了。1M上下文窗口加上原生多模态，要是真的开源Apache 2.0，国内一票闭源收费模型可以直接宣布倒闭。

Arthur_Simmons_Pro669

—

说真的用习惯了回不去了

Bryan.RobertsSr

—

做了三个月的开发者调研，结论就是：中小团队无脑选DeepSeek V3.2，便宜量大管饱。大企业的合规要求高的场景用Claude或GPT，毕竟数据安全这块DeepSeek确实有隐患。2025年初那次数据库泄露事件100多万行聊天记录和API密钥全暴露了，对企业用户来说这是硬伤。

Walter_ScottII

—

免费版居然偷偷把速率限制从60次降到10次每分钟了？没有任何通知直接改的，这操作真的很让人无语。

iMitchellTucker_2024

—

用DeepSeek写了一个月的代码了，感觉Python和Go的生成质量还不错，但React组件有时候会生成过时的API用法。整体来说配合Cursor用体验最好，纯网页端差一些。

DFlores_230

—

有没有人发现DeepSeek的错峰优惠真的很划算，凌晨用API价格直接打五折。我专门把批量任务调度到半夜跑，一个月省了好几百块。

Gerald702

—

用DeepSeek做小红书选题和文案效率真的高，一篇笔记从选题到写稿到配图建议十五分钟搞定，比我自己想半天强多了。不过内容质量还是要自己把控，AI写的有些太套路了需要手动润色。

何丽

—

DeepSeek中文能力确实比ChatGPT强太多了，古文翻译这块我对比过好几首诗，DeepSeek的意境把握明显更好。技术文档翻译也更准确。不过英文写作还是GPT和Claude的天下。

cAROL748

—

作为一个量化从业者，看到幻方2025年收益率56.6%真的很酸。DeepSeek三年烧了数亿美元完全靠母公司养着，这种模式国内找不出第二家。现在开始融资了说明梁文锋也扛不住了，毕竟V4的算力成本可不是闹着玩的。

John.RossZ

—

又宕机了！从昨天下午到现在快12个小时了完全用不了，这已经是这个月第三次了。做生产环境部署真的心惊胆战，谁能想到一个百亿估值的公司服务器稳定性做成这样。

JeanneAndrews

—

最近试了一下DeepSeek的Agent功能，V3.2的智能体能力确实有提升，工具调用准确率比之前高了不少。但是跟Claude的Computer Use比还是有差距，复杂的多步骤任务容易断链。

Kathleen.KellyZ7

—

我们实验室用DeepSeek跑了一批生物医学文献检索实验，发现幻觉率高达91%，就是会编造不存在的论文引用。同样任务GPT-4o的幻觉率只有39%左右。学术场景用DeepSeek一定要人工核实，千万别直接引用。

悠然457

—

Engram记忆模块那篇论文认真看了一遍，O(1)检索的思路很聪明，把静态知识编码成哈希索引避免重复计算。不过27B参数的演示模型效果虽然好，能不能scale到万亿参数还是未知数。

DennisRussellII

—

国产之光好吧

AlexanderMorales36919

—

DeepSeek在Reddit上被喷得最多的就是审查问题，确实有些政治话题完全没法讨论。但说实话作为开发者我主要用它写代码，跟审查有什么关系。API调用量大的话成本优势太明显了，这点小毛病完全可以接受。

蔡红

—

对比了一下DeepSeek V3.2和GPT-4o在同一套代码题上的表现，30道题DeepSeek对了25道，GPT-4o对了24道，基本持平。但价格差了将近十倍，性价比碾压。唯一的问题是DeepSeek偶尔会给出看起来正确但实际有细微bug的代码，需要仔细review。

smallzebra920

—

昨天测试了一下DeepSeek的金融数据分析能力，给它丢了一份财报让它分析，结果里面引用的营收数据全是错的，和实际数据差了15%左右。这个幻觉率在金融场景真的不能忍，用之前一定要自己核对数据。

IBfis

—

DeepSeek的MIT开源协议真的太良心了，商用完全没问题还可以二次开发。我们公司直接拿DeepSeek Coder做了个内部代码助手，效果比自己从头训练好多了。Llama虽然有开源但700MAU的限制太烦了。

翡翠879

—

刚知道微博和小红书都接入DeepSeek R1了，难怪最近搜索结果质量提升了不少。不过知乎直答接入的那版满血R1体验最好，回答比微博和微信AI搜索的都详细。

Jason_StewartIII0

—

我们团队现在的标准工作流就是：DeepSeek处理大批量的代码生成和数据处理，然后Claude做最终审核和面向客户的输出。两个模型配合着用成本降了70%质量也没怎么掉，真的很香。

DOkra

—

有人跟我一样2月份那次更新之后觉得DeepSeek变了吗？不再叫我名字了统一叫「用户」，推荐电影还来一句「够你看一阵子，不够再来要」——这爹味也太重了。文风变得又油腻又矫情，写出来的东西跟青春伤感文学似的。希望V4能把这毛病改了。

赵凤静

—

更新后变傻了吧，回答质量明显下降了好几个档次。业内人士说是为了V4做压力测试的极速版，牺牲质量换速度。要是这样我可以等，但别默认给我切到这个版本啊。

王杰雪

—

DeepSeek R1的推理链可视化真的很有意思，能看到它一步步怎么想的，有时候比最终答案还精彩。做数学题的时候会看到它先试了一种方法发现不行然后自己纠正换思路，这个自我纠错能力确实强。

JPerez369

—

免费版对于非开发者用户来说完全够用了，对话质量不输ChatGPT Plus的日常使用场景。唯一的缺点是没有多模态，图片识别还得用GPT-4o。等V4上了原生多模态估计会大一波。

孔雅

—

DeepSeek月活3.28亿这个数据太恐怖了，跟豆包形成双寡头格局。但月活高不等于最好用，程序员写代码还是更愿意用Claude或Cursor。不同场景选不同工具才是正解。

Megan_Hicks

—

试了一下把DeepSeek R1蒸馏到本地跑，用个70B的参数量在普通显卡上推理速度还可以，数学题准确率大概能到原始R1的85%左右。对不想用API但又想用国产模型的开发者来说是个好方案。

DeepSeek

In-depth Report

User Reviews

Similar Products