Compete for DeepSeek overflow traffic! Major applications should quickly connect. Users complain: they search for Douyin, but are recommended Video Number.

The advantages of open source and low cost have enabled DeepSeek to break through the high walls built by tech giants with computing power, and also have made domestic large model enterprises feel new competitive pressure. Currently, domestic Internet giants are seizing the opportunity to expand their layout with the popularity of DeepSeek. Following the launch of DeepSeek large models on cloud computing platforms owned by companies such as Alibaba, Tencent, ByteDance, and Baidu, applications have quickly followed suit. Baidu Search, Tencent Yuanbao, and other products have all integrated DeepSeek, attempting to take advantage of this wave of traffic and compete for market share. Applications are rushing to integrate and ride the wave of DeepSeek traffic. With the popularity of DeepSeek, AI user traffic brought by various major Internet companies has been highly valued, and various applications have successively integrated DeepSeek in an attempt to embrace the overflow of traffic from DeepSeek servers. For example, after integrating the DEEPSEEK-R1 full version, Baidu AI Search has been prominently placing advertisements for the platform. Alibaba's Dingding AI Assistant has fully integrated the DeepSeek series models, supporting deep thinking. The most high-profile integration is Tencent Yuanbao, which has been advertising on various platforms such as Douban, WeChat, and Zhihu about integrating Yuanbao with DeepSeek. Its marketing momentum is poised to surpass that of DouBao and Kimi. Intense advertising has brought users to Tencent Yuanbao. Yuanbao has already surpassed DouBao and is now ranked second on the mainland China's Apple App Store free list, with DeepSeek still ranking first. ByteDance has also chosen to embrace DeepSeek. After its Feishu platform announced integration with DeepSeek, ByteDance's Wukong browser has recently integrated the DeepSeek R1 model. Some dedicated DeepSeek users have told reporters from "Technology Innovation Board Daily" that after trying out several platforms, Tencent Yuanbao's response speed and logical answers are closest to the original version. "Some platforms, when used, are not as accurate as the official DeepSeek." Some users have found that after integrating DeepSeek into their apps, these Internet companies seem to have their own "calculations." "When I used Tencent's integrated R1 Yuanbao to search for content on the Douyin platform, it instead recommended content from the Video Number platform," one user said. Some users have also found that sometimes, the answers to the same question on the DeepSeek app and the DeepSeek full version integrated into the Baidu search box may differ. Especially when it comes to Baidu's own issues, the discrepancies in answers can be quite significant. In response, an industry expert in artificial intelligence said that since the DeepSeek model is fully open source, major Internet companies will definitely make some "tweaks" when integrating it into their products and services. It is not possible to help or promote competitors' products on their own DeepSeek application side. "Major Internet giants like Tencent, Baidu, and ByteDance have the technical ability to make adjustments to the open-source DeepSeek model, effectively adding their own small plug-ins to a public software." Cloud API doesn't make money, but it drives traffic. After experiencing a surge in user traffic, DeepSeek's official services are facing computing bottlenecks, and the API service recharge was temporarily suspended at one point. In the latest announcement, DeepSeek announced an off-peak discount activity for its API service from 00:30 to 08:30 daily, with DeepSeek-V3 reduced to 50% of the original price, and DeepSeek-R1 reduced to 25%, aiming to encourage users to use the service during off-peak hours. "Nowadays, using API for paid access often times out, so it should be guaranteed for paying users to use," complained a developer. According to sources, DeepSeek is one of the earliest companies in China to build a domestic WanCard NVIDIA cluster. "They have sufficient computing power for training, but the main bottleneck is GPU inference resources. They do not use third-party cloud services, and backend large-scale inference is still done using their own offline data centers," said an industry insider in cloud computing. In order to seize the traffic brought by DeepSeek, domestic cloud platforms have successively launched DeepSeek R1 APIs as official alternatives and provided discounts to attract users. For example, Alibaba Cloud offers users 1 million free tokens with Baolian, and additional fees are required beyond the free quota. DeepSeek full version R1 and V3 have a 50% discount and provide Alibaba Cloud vouchers. Large model fee-based APIs are not the main source of revenue for cloud services. Cloud providers value the additional product sales they bring. An Alibaba executive previously responded at a performance meeting, saying, "The revenue generated by charging APIs is still relatively small. Since DeepSeek is an open-source model, there will be a considerable number of developers who will develop vertical models and applications based on it, thereby using Alibaba Cloud. In addition, customers using the DeepSeek API will drive sales of other products on the cloud." Increasing investment in cloud and AI infrastructure. DeepSeek has sparked a round of investment in cloud and AI infrastructure among major domestic Internet companies. Alibaba Group CEO Daniel Zhang said that in the next three years, more than 380 billion yuan will be invested in cloud and AI hardware infrastructure. This investment exceeds the total amount invested by Alibaba Group in the past ten years. On February 25, Baidu announced the acquisition of YY Live, an online entertainment and live streaming business in China, from Huajun Group for approximately $2.1 billion. After the new agreement was reached, the approximately $1.6 billion previously deposited in the escrow account has been returned in full to Baidu. Baidu said it will invest the $1.6 billion in cloud and AI infrastructure development. Major Internet companies have not given up on competing with DeepSeek in the field of basic large models. Baidu will release the Wenzhen large model version 4.5 in mid-March. It is reported that the new version will significantly enhance the basic model capabilities, including multimodal and deep thinking capabilities. It was revealed at a previous financial report meeting that the Wenzhen 4.5 series is the most powerful large model in history, and it is planned to be officially open-sourced in June. In the past month, Alibaba has intensified...Multiple models have been released and open-sourced, including the first open-source of the visual generation model Wanneng 2.1. The number of derivatives of the Alibaba Cloud Qwen model has exceeded 100,000, making it the largest AI model family in the world.It has been reported that the DeepSeek-R2 model, originally scheduled to be released in May this year, is being accelerated in development, and may be released ahead of schedule. The technical iteration is still ongoing, and the outcome of the battle of large models is yet to be determined. Li Kaifu, CEO of Zero One Technology and Chairman of Innovation Works, believes that the future development of models will enter a watershed, mainly divided into two types: one is trillion or even tens of trillion level super-large models, and the other is billion-level large models. Billion-level models have the characteristics of practicality, speed, and lower costs, while super-large models are smarter and have higher ceilings. This article is reproduced from "Cai Lian She", GMTEight editor: Liu Xuan.