Why are universities the fastest to embrace DeepSeek?

Numerical Intelligence Frontline

follow with interest

2025-03-15

0 comments

34 views

0 favorites

19 minutes

  B-end product managers need to conduct pre-sales demonstrations, customized solutions, contract signing, etc., while C-end product managers need to plan activities, operate content, and motivate users, etc

After DeepSeek launched R1, universities became one of the fastest moving industries and also the earliest to challenge application scenarios and technical barriers.

DeepSeek threw a huge rock at the beginning of the year, causing ripples in various industries.

Multiple fields such as government affairs, finance, healthcare, and manufacturing are constantly evolving. Universities are no exception. Several industry insiders have revealed that in the wave of DeepSeek's large-scale model application boom at the beginning of 2025, universities are one of the most proactive and responsive industries.

Recently, many universities in China are intensively announcing the introduction of DeepSeek service. Dozens of schools have announced the introduction of DeepSeek, "observed Fan Chun, director of the System Management Office at Peking University's Computing Center. Not only universities, but also some high schools are introducing DeepSeek.

On March 13th, the Computing Center of Peking University also issued a notice stating that in addition to providing DeepSeek dialogue services based on public cloud for all faculty and students, the DeepSeek full blooded versions R1 and V3 deployed locally by Peking University have also been deeply adapted to teaching application scenarios, providing services to multiple artificial intelligence applications on campus, including Peking University Wenxue, AIMD (Peking University Artificial Intelligence Medical Doctor), Huaxiaobei, and financial AI teaching assistants, becoming another university to officially introduce DeepSeek services.

After DeepSeek came out, many business department heads on campus hoped to make their applications usable and have been urging us to launch DeepSeek Fan Chun said.

What is a B-end product manager? What is the difference between a C-end product manager and a C-end product manager?

In B2B product managers, B stands for Business, which means business. Firstly, B2B product managers need to understand the importance of this position and design product solutions that are more suitable for the project's needs. Ultimately, in their daily work, B2B product managers

View details>

What impact and changes has DeepSeek brought to the university market? What are the new opportunities and requirements for the landing of large-scale university models in the market in 2025?

Why are domestic universities embracing DeepSeek?

In the wave of DeepSeek application boom at the beginning of 2025, universities took active and rapid actions.

Universities and research institutes are considered to have responded earlier in various industries. Because DeepSeek was the first to explode in the academic circle, and then expanded to the technology application circle and industrial circle. "Wu Bingkun, founder and CEO of Zhongshu Xinke, a developer of large model industry applications, told Shuzhi Frontline that since the second day of the Lunar New Year, they have received many inquiries from clients, including many university clients, on how to use DeepSeek models combined with local data to build their own exclusive business scenarios.

Fan Chun, Director of the System Management Office at Peking University's Computing Center, also revealed that although Peking University has only recently officially launched DeepSeek related services, they have actually been paying attention to this matter for a long time and have started relevant preparation work, so that the model can be directly opened to the whole school for use and run stably after it is launched.

"Last year, many schools have continued to have demand for large models and are trying, but the explosion of DeepSeek at the beginning of this year has really further increased the demand." Zhang Xiangyue, senior solution architect of Baidu AI Cloud education industry, said that they have seen a lot of demand in campus management, scientific research and other directions.

Why has DeepSeek been able to generate such a large-scale and rapid follow-up trend in the field of universities?

On the one hand, this is because DeepSeek has significantly improved its basic model capabilities, filling the gap in the demand for localized deployment of large models that were previously difficult for universities to obtain a good user experience.

Industry observation shows that in many scenarios, closed source models were unable to achieve private delivery, and even if they could, the cost would be very high. Previously, although open-source models could also be deployed locally, they were not as thoroughly open sourced as DeepSeek. The highest versions were charged, and the cost was still high. The quality of localized deployment for smaller parameter scale models that were opened up was generally not high. "School teachers used this model and thought it was stupid, so they wouldn't use it anymore and would rather use external models. However, due to the high requirements of schools for data not to be out of domain, it cannot be deployed locally, which means that many applications related to large models cannot truly take off.

The emergence of DeepSeek has changed this situation, making more college users willing to use large models.

DeepSeek is the first open-source model that, after localized deployment, can achieve a level of intelligence comparable to the industry's best models provided by OpenAI Fan Chun told Shuzhi Frontline that many schools have previously developed AI applications that can now be deployed locally, and more teachers hope to develop new applications based on localized DeepSeek.

On the other hand, DeepSeek's higher cost-effectiveness also drives the enthusiasm of university users.

Fan Chun revealed that although the full blooded version of DeepSeek has a much larger number of parameters than some open-source models that can be localized and deployed before, and requires higher computational costs, the effect is also better. The return on investment is cost-effective, and people's willingness to use it is higher. Last year, Professor Zhao Hai from Shanghai Jiao Tong University decisively abandoned GPT-4 and switched to DeepSeeker V3 to generate synthetic data and develop large-scale vertical models. The reason is that DeepSeeker V3's performance is on the same level as GPT-4, but its price is only one tenth of it.

In addition to the full blooded version, the DeepSeek distilled version has a lower cost and can achieve good results in some scenarios, becoming a major choice for many college users.

For example, we can first do some DeepSeek model distillation to reduce computing power requirements. Based on this, we can then combine customer specific data to do some reinforcement learning to ensure better results. In theory, we can reduce the cost of AI projects to 10% of the original. Considering engineering and integration with other models of different sizes, the cost can reach about 1/4 of the original, "said Wu Bingkun from Zhongshu Information Science.

In addition, universities increased their investment in intelligent computing infrastructure last year. Fan Chun told Shuzhi Frontline that some universities' existing intelligent computing server resources can also quickly meet the computing power needs of this wave of DeepSeek localization deployment and achieve rapid deployment.

In fact, due to the surge in application enthusiasm and the increasing demand for computing power, universities are also important purchasers in the recent wave of all-in-one machine procurement. Li Chuan, General Manager of the Product Center of Shenzhou Digital Innovation Business Group, told Shuzhi Frontline that they see that the five industries with high demand for all-in-one machines are government affairs, education and research, medicine and healthcare, finance, and intelligent manufacturing.

The combination of model effects, private deployment, controllable costs, and computing resources has immediately released the scenario requirements of various universities in China, and the implementation process of large models in universities is being greatly accelerated.

The application of the 02 teaching and research management service has been fully launched, but there are still barriers to implementation

Driven by DeepSeek, the application scenarios of large models in the field of universities are rapidly expanding.

"By analogy to the mobile Internet era, 2024 is equivalent to the early days of the iPhone, and many apps are trying new things. After DeepSeek came out, it seems that we have directly stepped into the stage of iPhone 4 and iPhone 4S overnight." Wu Bingkun from Zhongshu Information Science told Shuzhi Frontline that in the past, many teachers only had a trial and error attitude towards large models, but after DeepSeek came out, everyone had strong confidence and motivation - 'I want to use it'.

Therefore, many universities that have recently announced their integration into DeepSeek have actually applied big models to various scenarios such as teaching, research, management, and services to try and explore.

Among them, AI for campus information management and service scenarios is the first to be implemented.

At the beginning of last year, we jointly developed the 'iHuadian' AI assistant with North China Electric Power University, which is supported by a large model. Campus cards, online fee recharge, library borrowing inquiries, and other related services can all be obtained with one sentence, "Baidu's Zhang Xiangyue told Shuzhi Frontline.

In the latest wave of DeepSeek applications, the vast majority of universities have also empowered this type of scenario as the first step towards implementation. Various question and answer based campus AI management assistants with relatively low barriers to entry are being widely replicated in more universities.

Various teaching aids are one of the scenarios that universities often explore and implement.

For example, Peking University recently announced the deep adaptation of several major artificial intelligence applications of DeepSeek's full blooded versions R1 and V3- Peking University Wenxue (AI assisted teaching system), AIMD (Peking University AI Doctor of Medicine), Huaxiaobei, Financial AI Assistant, etc., which are all AI applications that it has previously developed for teaching scenarios.

Last year, Zhongshu Information Science assisted some universities in upgrading their student learning effectiveness evaluation system and teacher teaching quality evaluation system through large-scale models. There is an impossible triangle in the education industry, which is to achieve large-scale, high-quality, and personalized solutions simultaneously. The emergence of AI is changing this situation.

In the college student training plan, there are a large number of evaluation dimensions. If traditional methods rely on programming to target tens of thousands or even tens of thousands of students, only some general evaluations can be conducted, making it difficult to achieve personalized customization. But with the large model, Zhongshu Information Science has achieved breakthroughs in classroom activities through the collaborative cooperation of multiple intelligent agents. Some intelligent agents are responsible for transcribing teaching speech, while others analyze and compare teaching levels; Some generate in class quizzes, while others grade papers on site; Furthermore, for each student's weak points, we recommend learning materials to achieve breakthroughs in personalized, large-scale, and high-quality teaching.

Some more complex teaching assistance scenarios that were previously difficult to implement are also being explored at an accelerated pace.

For example, in interactive teaching scenarios, as a teacher has to face many students, it is difficult to design teaching plans based on individual characteristics. Therefore, some universities hope to use intelligent AI teaching assistants to label each student's learning behavior and design targeted teaching plans to achieve personalized teaching.

Previously, there was still a need for extensive optimization at the model layer, as well as a lot of intelligent agent orchestration, to create an AI interactive course with high customization costs. "Baidu's Zhang Xiangyue told Shuzhi Frontline that this year, with the improvement of model capabilities and the continuous reduction of technical barriers, related costs are expected to be amortized.

In addition, teaching and practical training are also the demand points of universities. After the big model becomes popular in 2023, some universities will offer practical courses that require students to do some hands-on exercises, "Luo Lei, the head of intelligent computing products at iFlytek's Spark Corps, told Shuzhi Frontline. The demand for this scenario has further increased this year, which in turn has increased the demand for all-in-one machines." One classroom can be equipped with one all-in-one machine.

Wu Bingkun also stated that previously, the worst software and hardware for this AI training platform would cost two to three million yuan. After DeepSeek arrived, due to open source and reduced computing power, "it may suddenly reach the order of tens of thousands", and application scenarios will also be opened up.

In addition, in terms of scientific research, the trend of AI for Science has been on the rise since the second half of last year. For example, Shanghai Jiao Tong University has transformed its traditional research model and hopes for a closer integration between science and AI. They cooperated with Baidu AI Cloud to build an AI for Science scientific data open source platform to support the training of Magnolia Magnolia scientific model. Relying on the AI for Science platform, Shanghai Jiao Tong University has published the scientific achievements of AI+cities on the cover of Nature Computational Science.

However, the industry also sees that although the scenarios for universities to implement large models are rapidly expanding, the threshold for implementing large models is still not low.

Several industry insiders have told Shuzhi Frontline that the way to call from the cloud is relatively simple, but more universities still choose to deploy privately due to data security needs. And in this process, there are a large number of technical barriers to entry.

Firstly, you need to have equipment, and secondly, you need to have people. Either you have your own technical team, or you need to find a powerful service provider to do the corresponding development and optimization adjustment work, in order to support DeepSeek's stable service and prevent it from collapsing with too many people. "Fan Chun from Peking University told Shuzhi Frontline that everyone hopes to achieve better results with less resources and funding, but this also puts higher demands on universities and service providers. Therefore, currently, the effects of introducing DeepSeek in different schools may vary.

Now, universities and service providers are exploring some new methods. For example, on March 9th, Peking University and Huawei jointly released a DeepSeek full stack open-source inference solution to help developers improve inference performance.

Wu Bingkun from Zhongshu Information Science introduced that the open-source model is first distilled to obtain a smaller parameter model to reduce computing power requirements, and then the data is trained on the model through reinforcement learning instead of the previous fine-tuning and RAG. After reinforcement learning, the model can guide its deep inference quality in reverse. Through this method, some users have tested that the accuracy is 15% to 20% higher than using DeepSeek directly.

03 Mid to long tail demand is on the rise

As DeepSeek launches a wave of applications, the industry observes that the mid to long tail demand in the university field is on the rise.

"In the past, teachers in some schools or vocational schools in both Africa and China had relatively little money at their disposal. After DeepSeek came out, first, its computing power was greatly reduced. Second, the model itself did not need money, and such customers may be able to use it in the future." Wu Bingkun from Zhongshu Information Science said.

Shenzhou Digital's Li Chuan also told Shuzhi Frontline that they have seen demand not only from top universities, but also from vocational colleges and universities. Moreover, there is a need for localized deployment of large models not only at the school level, but also at the department level, and even at the teaching and research group level.

Fan Chun from Peking University observed that this year, big models are shifting from assisting research to more assisting in improving services. "Many non double major universities find it difficult to make research decisions because the investment in AI is relatively large. The president will think, 'After I invest, can I produce results or effective results? Because research results are inherently random, but services can be more easily seen with results.'

In response to these characteristics, some service providers in the industry chain have already laid out.

Some enterprises, especially cloud companies, are strengthening their full stack advantages. For example, from the external information of Baidu AI Cloud, there are domestic independent Kunlun Core ten thousand card cluster, the computing power scheduling center of Baige heterogeneous computing power platform, the PaddlePaddle deep learning framework and Qianfan big model platform, as well as the measures of industry university research collaborative innovation.

On March 3rd, iFlytek teamed up with Huawei to iterate the new generation of Spark all-in-one computers and launched for the first time an all-in-one product targeting five new application scenarios: healthcare, higher education, government affairs, law enforcement, and law.

Universities will still be a key market for Zhongshu Information Science this year. On February 13th, Zhongshu Xinke also partnered with China Mobile to launch a large model all-in-one machine, pre installed with DeepSeek and other large models, and directly set the rental price to thousands of yuan per month through the "rental purchasing" method. Many customers now have strong motivation for AI intelligent agents, but they can't wait to initiate projects and go through the process, "said Wu Bingkun. This model can meet customers' needs for quick experimentation.

Shenzhou Digital's Li Chuan revealed that they have launched a large model all-in-one machine called "Family Bucket", which can provide solutions ranging from over one million yuan to over 100000 yuan. Some teaching and research groups or individual graduate students in schools can choose lightweight solutions to implement DeepSeek deployment.

Last year, many schools had already set a good example, and with the release of DeepSeek, the threshold for use was lowered. This year, schools with long tail designs can avoid many detours by replicating it. Moreover, many schools built their computing power last year, and with the release of DeepSeek this year, they can use all the computing power from last year. Therefore, this year is definitely a year of explosive application growth, "said Zhang Xiangyue from Baidu.

Fan Chun believes that with the further application of large-scale models, universities will continue to increase their investment in computing infrastructure this year. On the one hand, the trend of AI for Science is still ongoing and requires more computing power. On the other hand, the application of AI in more campus scenarios such as teaching, management, and services will also drive further growth in computing power demand.

We will continue to observe how the implementation of large-scale university models will develop.

Article | Edited by Zhou Xiangyue | Zhao Yanqiu

This article is written by everyone who is a product manager [Digital Intelligence Frontline]. WeChat official account: [Digital Intelligence Frontline]. It is original/authorized to be published by everyone who is a product manager. Reproduction without permission is prohibited.

The title image is from Unsplash, based on the CC0 protocol.

高校，为何最快拥抱DeepSeek？

数智前线关注

2025-03-15

0 评论34 浏览0 收藏19 分钟

DeepSeek推出R1后，高校是动作最为迅速的行业之一，也是最早挑战应用场景和落地技术门槛的行业。

DeepSeek开年扔下的一块巨石，砸起了各行各业的水波荡漾。

政务、金融、医疗、制造等多个领域，动态不断。高校也不例外，多位行业人士透露，在2025年初的这波DeepSeek掀起的大模型应用热潮中，高校是态度最为积极、动作最为迅速的几大行业之一。

近期，国内多所高校，正在紧锣密鼓官宣引入DeepSeek服务的消息。“有几十个学校都宣布引入了。”北京大学计算中心系统管理室主任樊春观察，不仅是高校，甚至一些中学也在引入DeepSeek。

北京大学计算中心也在3月13日发布通知，除了为全校师生提供基于公有云的DeepSeek对话服务，北大本地化部署的DeepSeek满血版R1和V3也已深度适配教学应用场景，向校内多项人工智能应用北大问学、AIMD(北大人工智能医学博士）、化小北、金融AI助教提供服务，成为又一家官宣引入DeepSeek服务的高校。

“DeepSeek出来后，校内很多业务部门的相关负责人希望能让他们的应用也使用上，一直在催着我们上线DeepSeek。”樊春说。

什么是B端产品经理？和C端产品经理有什么区别？

B端产品经理中的B是Business，商业的意思，B端产品经理首先就要理解这个职位的重要性，要设计出更适合这个项目需求的产品方案，B最终产品经理在日常工作中...

查看详情 >

DeepSeek到底给高校市场带来了哪些影响和改变？2025年，高校大模型落地市场，又有哪些新机遇和新要求？

01 国内高校，为何纷纷拥抱DeepSeek?

2025年初的这波DeepSeek应用热潮中，高校的动作积极且迅速。

“高校和科研院所算是各行各业里响应比较早的。因为DeepSeek是最先在学术圈‘爆炸’，然后扩大到技术应用圈和产业圈。”大模型行业应用开发商众数信科创始人兼CEO吴炳坤告诉数智前线，他们从大年初二开始，就陆陆续续接到很多客户的咨询，全是问怎么用DeepSeek模型结合本地数据，构建自己的专属业务场景的，这其中就包括不少高校客户。

北京大学计算中心系统管理室主任樊春也透露，虽然北大最近才正式上线DeepSeek相关服务，但他们其实很早就关注到这个事情，开始了相关准备工作，以便模型上线后能直接敞开给全校使用，且稳定运行。

“去年，有不少学校都经陆陆续续有大模型方面的需求，在去做尝试，但今年初DeepSeek的爆火确实进一步加大了需求。”百度智能云教育行业高级解决方案架构师张湘悦说，目前他们在校园管理、科研等方向都已看到不少需求。

为什么DeepSeek能在高校领域掀起如此大规模且迅速的跟进潮？

一方面是因为DeepSeek在基础模型能力上的大幅提升，填补了此前高校难以得到好用的本地化部署大模型使用体验的需求空白。

业界观察，之前闭源模型在很多场景下，无法做到私有化交付，即便能做到，成本也很高昂。而以前的开源模型，虽然也可以本地化部署，但不像DeepSeek开源这么彻底，最高版本都是收费的，成本仍然较高，开放出来的更小参数规模的模型，本地化部署的质量也普遍不高，“学校的老师用了这个模型觉得很笨，就不会再用，宁肯用外面的模型”。但由于学校也对数据不出域有着较多要求，不能本地化部署，也就意味着大模型相关的很多应用无法真正起来。

DeepSeek的出现改变了这种局面，让更多高校用户愿意去将大模型用起来。

“DeepSeek是第一个本地化部署后，能达到与OpenAI提供的业界最好的模型差不多智能水平的开源模型。”樊春告诉数智前线，很多学校之前做了一些AI应用，现在能够本地化部署了，更多老师希望基于本地化DeepSeek去做新应用开发。

另一方面，DeepSeek在成本上实现的更高的性价比，也催动着高校用户们的热情。

樊春透露，虽然满血版的DeepSeek参数量要比之前能够本地化部署的一些开源模型大上很多，算力方面所需的成本更高，但效果也更好，投资回报是划算的，大家的使用意愿更高。而上海交大赵海教授去年果断弃用GPT-4，改用DeepSeek-V3生成合成数据，开发垂类大模型，原因正是DeepSeek-V3性能与GPT-4处于同一量级，而价格只有其十分之一。

而满血版之外，DeepSeek 蒸馏版的成本更低，也能在一些场景实现不错的效果，成为不少高校用户的一大选择。

“比如我们先去做一些DeepSeek的模型蒸馏，来降低算力需求，在此基础上，再结合客户独有的数据，去做一些强化学习，保证更好的效果，理论上，能将AI课题成本降低到原来的10%，考虑工程化及与其他大小模型的融合，成本大概能达到是原来的1/4。”众数信科吴炳坤说。

另外，高校去年曾进行过一波智算基础设施投入的加码。樊春告诉数智前线，一些高校手上已有的智算服务器资源，也能更快满足这波DeepSeek本地化部署的算力需求，实现快速上线。

事实上，由于应用热情的激增，带动着算力使用需求的增加，在最近的一波一体机采购潮中，高校也是重要的采购方。神州数码信创业务集团产品中心总经理李川告诉数智前线，他们看到目前对一体机需求较多的5大行业分别是政府政务、教育科研、医药医疗、金融和智能制造。

模型效果、私有化部署、成本可控、算力资源等几个条件叠加在一起，国内各高校的场景需求，一下就释放出来了，大模型在高校的落地进程正被大幅加速。

02 教研管服应用全线开启，但落地仍有门槛

DeepSeek带动下，高校领域的大模型应用场景正在快速扩大。

“类比移动互联网时代，2024年相当于当年iPhone早期，很多App做尝鲜。DeepSeek出来后，一夜之间，大家好像直接跨到了iPhone4、iPhone4s的阶段。”众数信科吴炳坤告诉数智前线，以前，很多老师对大模型只是抱着试试看的态度，DeepSeek出来后，则变成了大家都有强烈的信心和动力——“我要把它用起来”。

也因此，最近官宣接入DeepSeek的不少高校，实际都已将大模型用到了教学、科研、管理、服务等全线场景中去尝试、探索。

其中，校园信息化管理和服务场景的AI是最先落地的。

“去年初，我们跟华北电力大学联合开发了大模型支撑的‘i华电’AI助理，校园卡、网费充值、图书馆借阅查询等，都可以一句话，获取相关服务。”百度张湘悦告诉数智前线。

而在最新这波DeepSeek应用潮中，绝大部分高校也都将这一类场景的赋能，作为落地第一步，各种门槛相对较低的问答式校园AI管理助手，在更多高校中进行大规模复制。

各种教学辅助环节，则是高校较多去探索落地的场景之一。

比如北大近期宣布深度适配了DeepSeek满血版R1和V3的几大主要人工智能应用——北大问学（AI辅助教学系统）、AIMD（北大人工智能医学博士）、化小北、金融AI助教等，就都是其此前已经开发出的一些面向教学场景的AI应用。

众数信科则在去年，通过大模型助力一些高校实现了学生学习效果测评系统和教师教学质量测评系统的升级。教育行业存在不可能的三角，即同时实现大规模、高质量和个性化。而AI的出现，正在改变这种现状。

在高校学生培养计划中，存在大量评测维度，如果按照传统做法，依靠编程，针对上万乃至数万学生，则只能进行一些通用评测，很难实现个性化定制。但有了大模型后，众数信科针对课堂环节，通过多个智能体协同合作，实现了突破。如有的智能体负责转录授课语音，有的进行授课水平的分析对比；有的生成随堂小测试卷，有的现场批改试卷；还有的针对每个学生的薄弱点，进一步推荐学习材料，在个性化、规模化和高质量教学上实现突破。

一些此前难以实现的更为复杂的教学辅助场景也在加速被探索。

比如互动式教学场景，由于一个老师要面对很多学生，很难根据个体特性去设计教学方案，一些高校因此希望能通过智能AI助教，给每位学生的学习行为打标签，并设计针对性的教学方案，实现因材施教。

“之前还需要在模型层做大量优化，以及很多的智能体编排，做一门AI互动课，定制化成本很高。”百度张湘悦告诉数智前线，今年随着模型能力的提升，技术门槛不断降低，相关成本有望被摊薄。

除此之外，教学实训，也是高校的需求点所在。“从2023年大模型火起来后，一些高校就会开一些实践课程，需要学生去做一些实操。”科大讯飞星火军团智算产品负责人罗雷告诉数智前线，这种场景的需求今年进一步增长，并进而增加了对一体机的需求，“一间教室可以配一台一体机”。

吴炳坤也表示，此前这种AI实训平台，最差的软硬件加在一起也要两三百万元，DeepSeek来了后，由于开源加上算力的降低，“可能一下子就拉到几十万的一个数量级”，应用场景也随之打开。

另外，在科研方面，AI for Science的热潮自去年下半年就已经起来，像上海交通大学，已转变传统科研模式，期望科学与AI更紧密的结合。他们与百度智能云合作建成了AI for Science科学数据开源开放平台，支撑白玉兰科学大模型的训练。依托AI for Science平台，上海交大已在Nature Computational Science封面，发表了AI+城市的科学成果。

不过，业界也看到，虽然高校落地大模型的场景在快速拓宽，但大模型的落地门槛仍然不低。

多位行业人士告诉数智前线，云端调用的方式相对简单，但更多高校出于数据安全需求，还是选择私有化部署。而这个过程中，有着大量的落地技术门槛。

“首先你得有设备，其次还得有人，要么自己有技术团队，要么找有实力的服务商，做相应的开发和优化调整工作，才能支撑DeepSeek稳定服务，不至于人一多就垮掉。”北京大学樊春告诉数智前线，大家都希望能用更少的资源和经费，达到一个更好的效果，但这也对高校和服务商们提出了更高的要求。也因此，目前，各个学校引入DeepSeek后的效果 ，可能都是不一样的。

现在，高校和服务商们在探索一些新方法。比如，3月9日，北京大学和华为联合发布了一套DeepSeek全栈开源推理方案，以此来帮助开发者提升推理性能。

众数信科吴炳坤则介绍，先将开源模型蒸馏，获得更小参数的模型，来降低算力需求，再通过强化学习而不是之前的微调和RAG，将数据训练到模型上去。而模型经过强化学习以后，可以反向指导它的深度推理质量。通过这种方法，一些用户实测将比DeepSeek拿来后直接使用，准确率高出15%~20%。

03 中长尾需求在起来

随着DeepSeek卷起应用浪潮，业界观察，高校领域的中长尾需求正在起来。

“以前一些双非院校或职校里的老师，平时能够掌握的经费是比较少的。DeepSeek出来后，第一，算力大大降低了，第二，模型本身也不需要钱了，这类客户未来可能就都能用得上了。”众数信科吴炳坤说。

神州数码李川也告诉数智前线，现在不光是头部院校，他们在专科、职教都已经看到了需求，而且不只在学校级别，在院系级，甚至教研组级别，也都有大模型本地化部署的一些需求。

北京大学樊春则观察，今年，大模型正在从助力研究，进入更多去助力改善服务，“很多双非院校做研究决策比较困难，因为AI的投入都比较大，校长他会想，我投入后，能不能出成果或者有效果，因为研究出成果本来就有随机性，但服务是可以比较容易看到效果的。”

针对这些特点，一些产业链上的服务商已经布局。

一些企业，尤其是云公司，强化全栈优势。如从百度智能云对外的信息看，有国产自主昆仑芯万卡集群、百舸异构算力平台的算力调度中枢、飞桨深度学习框架和千帆大模型平台，以及产学研协同创新的举措。

科大讯飞在3月3日联合华为，迭代新一代星火一体机，并首次推出针对医疗、高教、政务、警务、法律五大全新应用场景的一体机产品。

高校仍将是众数信科今年的一大重点市场。2月13日，众数信科还联合中国移动，推出了大模型一体机，预装DeepSeek等大模型，并通过“以租代购”的方式，将租用价格直接打到了每月千元档位。“现在很多客户对于AI智能体有极强的动力，但等不及去立项、走流程。”吴炳坤说，这个模式可以满足客户快速尝鲜的需求。

神州数码李川透露，他们推出了大模型一体机“全家桶”，大到超百万元，小到10余万元级别的方案都可以提供，学校的一些教研组或个别研究生，都可以选用轻量级的方案，去实现DeepSeek部署。

“去年很多学校已经都打好样了，而且DeepSeek一出来，也降低了使用门槛，今年中长尾的学校，再去复制，就能少走很多弯路。而且很多学校去年都建设了算力，今年DeepSeek出来，正好把去年的算力都能用上，所以今年肯定是应用井喷的一年。”百度张湘悦说。

樊春则判断，随着对大模型的进一步应用，高校今年的算力基础设施投入也会继续加大，一方面，AI for Science的风潮还在继续，需要较多算力，另一方面，AI在教学、管理、服务等更多校园场景的应用，也会带动算力需求进一步增长。

高校大模型落地将如何发展，我们还将继续观察。

文｜周享玥‍‍‍ 编｜赵艳秋‍‍

本文由人人都是产品经理作者【数智前线】，微信公众号：【数智前线】，原创/授权 发布于人人都是产品经理，未经许可，禁止转载。

题图来自Unsplash，基于 CC0 协议。