As DeepSeek continues to be popular, people are not satisfied with the use of DeepSeek on the web and APP side, and try to localize DeepSeek. Localisation involves installing DeepSeek's large AI models on local computers, without relying on networks or cloud services. Reporters searched video websites and found that many users uploaded tutorials on how to deploy DeepSeek to local computers, and many videos were viewed more than 1 million times.
DeepSeek sparks local deployment boom
Teaching people to deploy DeepSeek has also become a business. Through searching for "DeepSeek local deployment" on the e-commerce platform, the reporter found that many shops have opened up DeepSeek local deployment business, and the unit price of these services ranges from a few dollars to dozens of dollars, and some of these services have recently been purchased by 1,000 people.
An AI enthusiast who has tried to deploy told reporters that the response speed of the network side is slow, and when the traffic is too large, there is often "the server is busy, please try again later." In order to get a better experience, he tried to use DeepSeek for local deployment. It is understood that local deployment does not need to master deep programming knowledge, through the tutorial step by step, you can deploy successfully.
Zhang Yi, chief analyst at IIMedia Consulting, told reporters: "Local deployment supports individuals to make some customized modifications to DeepSeek according to their needs, which is also one of the driving forces." Zhang Yi added that personal data in the local deployment does not go to the cloud, which can meet the privacy needs.
DeepSeek published models with varying numbers of parameters, from as small as 1 billion parameters to as large as 671 billion parameters, and the larger the parameters, the greater the computational resources required. Due to the limited computing resources of devices such as personal computers and mobile phones, the 671 billion parameter DeepSeek model often cannot be deployed locally. "A typical laptop can only deploy a billion-parameter version, but a PC with a good GPU or high memory (say 32GB) can run a 7-billion-parameter version of DeepSeek." The AI technology enthusiasts told reporters.
In terms of the effect of local deployment, the smaller the version of parameters, the worse the answer quality of the large model. "I tried the 7 billion parameter version of DeepSeek deployed locally, and it ran smoothly, but the answer quality was much worse than the cloud version, and the effect of the smaller parameter version was even worse." The above AI enthusiasts said.
Under the heat of DeepSeek's local deployment, AI PCS that specifically add NPU to PCS are expected to usher in sales growth. It is understood that Huawei, Lenovo and other computer brands have launched AI PC, this new PC is equipped with specialized processing of local deployment of AI large model computing processor chips, these processor chips are provided by Intel, AMD, Qualcomm and other chip factories.
These AI PCS can be deployed locally and run smoothly tens of billions of parameters of the AI large model, such as this CES 2025, AMD launched Ryzen AI max series processors, saying that the computer can run 70 billion parameters of the AI large model. However, the AI PC equipped with the processor chip is expensive, and it is understood that the price of an Asus game book is nearly 15,000 yuan. In addition, some people have questioned that spending a lot of money to buy AI PCS, carry out local deployment of AI large models, and achieve functions that are highly overlapping with cloud AI large models, AI PC is just a gimmick of manufacturers.
Enterprises try to deploy DeepSeek locally
In addition to individuals opening DeepSeek local deployment, enterprises are also beginning to try. On the first day of the working day of the Year of the Snake, Wang Jiahui, the founder of Timwei Ao, released a wechat circle of friends: "DeepSeek large model local computer deployment experience success, import coal mine knowledge safety database for questions and answers, the next step is to combine it with industrial field operations."
Timviao is a company that provides industrial management solutions for the mining industry, oil industry and other industries, using AR glasses and AI software to provide real-time observation and guidance for the maintenance, quality inspection and troubleshooting of industrial field personnel. Wang Jiahui told reporters that only Tongyi Qian ask AI large model to create a local knowledge base question and answer, in view of DeepSeek better reasoning ability, he is considering DeepSeek and business deep integration.
"On the basis of DeepSeek, we fine-tune specific parameters or secondary develop them to fit IT systems and implement new functions based on the needs and data of specific industrial scenarios." "Our goal is to deploy DeepSeek locally and interact with cameras in the field to better identify dangerous operations on site, and to implement functions such as hidden danger detection and product quality inspection," Wang told reporters.
He believes that whether industrial customers adopt on-premises deployment depends mainly on the confidentiality of the data. "Some state-owned enterprises, military and medical equipment companies often ask us to implement local deployment solutions because they have high requirements for data security." He further added: "Non-secret scenarios can use cloud access solutions, although there will be operational delays, but the impact is small, and the solution price is lower."
For on-premises deployments, these customers require servers equipped with 4-card or 8-card Gpus to implement DeepSeek local inference services. "My customers generally choose Nvidia's consumer graphics cards to configure their servers, such as the 4090, for better value for money." "If customers have domestic configuration requirements, we will buy more expensive domestic GPU graphics cards," Wang said.
In addition to industry, more and more enterprises are starting to deploy DeepSeek locally. Sinolink Securities said that DeepSeek can be applied to information retrieval, document processing, industry research, market research and other scenarios. In addition, enterprises in the medical industry, network security and other industries have also recently deployed DeepSeek locally, including Wanda Information and Qihoo 360.
Zhang Yi told reporters that as enterprises expand their requirements for localized deployment, the demand for domestic reasoning computing power will increase, and the United States bans high-end chips, domestic chip computing power companies will usher in greater opportunities.
AI applications will explode
Qualcomm CEO Cristiano Amon said that DeepSeek-R1 is a turning point for the AI industry, AI reasoning will migrate to the end side, AI will become smaller, more efficient, and more customized, and AI large models and AI applications based on specific scenarios will appear. China Aviation Securities Research report believes that DeepSeek-R1 shows that the end-to-end AI deployment will become more inclusive, and the era of all things intelligent will accelerate.
Open source will entice more developers to build apps on top of DeepSeek. Domestic GPU cards such as Huawei Centeng, Moore Thread, Bishi Technology, and Daywise Core have been adapted to DeepSeek; Tencent Cloud, Alibaba Cloud, mobile cloud, Huawei Cloud and other cloud manufacturers also completed the adaptation with DeepSeek. The adaptation optimization of domestic computing power is expected to further reduce the cost of inference side.
As the domestic app payment habit is not yet fully mature, the commercialization of AI applications may be hindered. Meng Can, chief of National Finance Computer, believes that the United States has a 10-year or even 20-year basis for application payment, which is helpful for the commercialization of AI applications, while the domestic market will be slow due to the lack of such a foundation. However, the country is constantly catching up, and the time schedule is expected to be reduced to less than half a year.
Your message must be between 20-3,000 characters!
