GuShiio.com Article source: Hedgehog Commune
Image source: Generated by AI
Do you pick up DeepSeek? It is becoming a new challenge for the decision-making levels of large factories such as Tencent, Baidu, Ali, and Byte.
In Houchang Village, Haidian Huangzhuang, Future Science and Technology City, Binhai Building, TIT Creative Park… countless questions about DeepSeek were raised, and some route-related decisions were made. Soon, these decisions were transmitted to users:
On February 15, WeChat officially tested the new “AI Search” function and connected to DeepSeek-R1 ‘s “Deep Thinking” service, which users can use for free; on the evening of February 16, Baidu search officials announced that they would fully access DeepSeek. Earlier, on February 12, Ali’s office software nails were fully connected to the DeepSeek series models; on February 4, ByteDance’s volcanic engine announced that it fully supports the DeepSeek series models; in the evening of the same day, Jingdong Cloud officially launched the DeepSeek-R1 and DeepSeek-V3 models.
It seems that all major manufacturers have made the same decision,But carefully distinguish between “connect” and “connect” reflects the different choices of BAT at this moment.
Will the business-side products be picked up together with our own large model products? Will cloud business or main business be connected? At least so far, big manufacturers have not generated standard answers.
Tencent is the most decisive and byte is the most confident
Discovering the current ways in which major Internet companies access DeepSeek can be roughly divided into three categories according to the degree:
- Full access, including self-developed large model products
- Main business side product access
- Cloud service access
1) Full line access
Tencent is a representative of full-line access to DeepSeek.
On the evening of February 15, WeChat officially tested the new “AI Search” function. When using the “AI Search” function in WeChat, users who have been grayed out can choose the “Deep Thinking” mode, which provides “a more comprehensive answer provided by DeepSeek-R1 model after long thought.”
According to the current experience, the WeChat version of “Teacher D” only provides a single question and answer session and does not have the ability to have continuous dialogue. It does meet the definition of “AI search” and is a lightweight AI search tool rather than an AI assistant.
In terms of generation results, WeChat AI search connected to DeepSeek has the networking function enabled by default. The information source comes from the entire network. Content search and content generation within the WeChat ecosystem are not supported for the time being.
On February 16, the QQ browser officially announced that it would connect to DeepSeek-R1, supporting in-depth thinking, online search, multiple rounds of conversations, historical tracing and other functions. After evaluation by Hedgehog Commune, it was found that the fixed entrance on the mobile phone is under the search icon in the search box on the home page, and a large banner entrance is also displayed on the home page.
On February 17, Tencent Docs also posted a tweet announcing that it would “officially access the full blood version of DeepSeek-R1.” Users can choose to use Tencent’s self-developed mixed model or DeepSeek-R1 on mobile and PC.
Compared with the access to business-side products such as WeChat, Tencent Docs, and QQ browser, Tencent’s other decision may be more indicative of its determination to embrace DeepSeek-R1. On February 13, Tencent’s self-developed AI assistant “Tencent Yuanbao” was officially connected to DeepSeek-R1, allowing users to switch between using the Mixed Origin Model and DeepSeek R1.
Previously, Tencent Yuanbao was positioned as a “C-side AI assistant App based on the self-developed mixed model”, which is equivalent to Wen Xiaoyan to Baidu, Doubao to Byte, and Tongyi to Ali.
As of February 17, Tencent Yuanbao is the only player among the above-mentioned large-scale AI assistant products that actually connects to DeepSeek.
Will this choice have an impact on the self-developed Mixed Origin Model? Judging from the current product description, Tencent hopes that DeepSeek-R1 will provide users with more comprehensive answers after long thinking, while the Mixed Origin Model can provide users with more common and efficient answers, which is equivalent to asking users ‘questions about scenarios and problem complexity. The degree is segmented and two different options are provided.
If Tencent Yuanbao is regarded as a single product, this choice will undoubtedly greatly optimize the user experience; however, placed in the context of the AI research and development process of a big factory, and connected to an external open source model, the brand reputation and investment legitimacy of the self-developed model will pose challenges., and even future development confidence may ultimately require management to make a decision.
Therefore, in addition to Tencent, the AI assistants developed by several other major manufacturers have not yet been truly connected to DeepSeek.
2) Main business side product access
The access of companies such as Baidu and Alibaba to DeepSeek is mainly reflected in their main business-side products.
Previously, it was reported that “Wen Xiaoyan has connected to the DeepSeek-R1 model,” but according to the update log, Wen Xiaoyan only optimized the “photo problem solving function” through the DeepSeek-R1 model.According to the test conducted by Hedgehog Commune on February 17, Wen Xiaoyan currently has only three dialogue models that can be switched between: Wenxin 3.5, Wenxin 4.0 and Wenxin 4.0 Turbo.There is no option for DeepSeek-R1, and the “Photo Solving Assistant” is just one of the many intelligent agents built in Wen Xiaoyan.
In the announcement released on February 13, in addition to mentioning that Wen Xiaoyan will be “fully free from 0:00 on April 1”, it also mentioned that “the new in-depth search function will be launched in the near future.” From the subsequent statement, the in-depth search function is self-developed by the Wenxin Model. In the future, whether Baidu will fully access DeepSeek in its self-developed AI assistant products like Tencent has not yet been announced.
In contrast, the news of Baidu’s search connection to DeepSeek settled on the evening of February 16. Baidu Search issued an official announcement saying: “In order to enrich and more diversified search experience, Baidu Search will fully access the latest in-depth search functions of DeepSeek and Wenxin Model.”
“Will” means a state that has been determined but not completed. At 0:00 on February 18, Baidu search access to DeepSeek finally became complete. After searching for any keywords from the homepage of Baidu APP, select AI+ to enter the AI search page. The DeepSeek-R1 entry will appear as a label above the input box.
In addition to search services, Baidu Library and Baidu NetDisk are both services with many application scenarios for AI capabilities. Whether they will subsequently access DeepSeek is also worthy of industry attention.
Ali’s current choice is similar to Baidu.
As the main C-end product of Tongyi Large Model, Tongyi APP, which is positioned as an “all-round AI assistant”, has not yet been connected to DeepSeek. A number of Alibaba’s business-side products have been connected or announced to be connected to DeepSeek, such as office software Nail, which has supported users to use DeepSeek series models to create nail AI assistants, customize intelligent multidimensional forms, etc.;1688 announced that it will develop a business model based on DeepSeek and launch free digital employees.
Compared with Tencent, both Baidu and Ali have chosen to equip business-side products with DeepSeek’s basic capabilities first, while reserving some independent space for their self-developed large model products.
3) Cloud platform access
Before DeepSeek swept the “Spring Festival”, ByteDance’s bean buns were once DAU’s highest large-scale independent app. But after the Spring Festival, everyone discovered that the seemingly unique DAU advantage may be disintegrated in an instant.
On the issue of whether to connect to DeepSeek, Byte made the most “conservative” choice.
As of February 17, among the byte series products, only Volcano Engine and its products + Flying Book officially announced that the DeepSeek series models were launched. However, launching different models is a routine operation for cloud platforms. Well-known cloud platforms such as Baidu Intelligent Cloud, Huawei Cloud, Alibaba Cloud, and Tencent Cloud all announced the launch of the DeepSeek model in early February. Take Volcano Ark, a large model service platform owned by Volcano Engine, as an example. The large language models on its shelves include not only the large model of the bean bun series developed by Byte, but also DeepSeek, the GLM model of Intelligent Intelligence AI, and the Moonshot of the Dark Side of the Moon.
In addition to the “basic operations” of the cloud service platform, among the byte-based business-side products, only Flybook has the tendency to connect to DeepSeek, and the entrance is deeply hidden. Users need to create a new multidimensional table-add columns-search field shortcut before they can find the entry for DeepSeek R1.
The most “conservative” may also mean the most “confident”–Byte believes that in the near future (or without affecting the fundamentals of the business), he can bring his self-developed model to a level close to that of DeepSeek.
Faced with the “emergence” of users brought by DeepSeek, major Internet companies that were once at the same pace on the road of large-scale model exploration have embarked on different paths.
Connecting to DeepSeek does not mean All in DeepSeek
As the news of “XXX Connected to DeepSeek” flew like snowflakes, many people began to use “All in DeepSeek” to describe the development of AI products during this period.
But connecting to DeepSeek does not mean All in DeepSeek.
Tencent chose to access DeepSeek across the board, taking the present as the node, regardless of sunk costs, and accessing the external model with the greatest current user demand, the best effect, and the best response. At present, this choice is DeepSeek. If in the future, Dark Side of the Moon or Intelligent Spectrum AI develops an open source in-depth thinking model with better results and greater user needs, Tencent may also replace it.
Alibaba and Baidu chose to access the main business side because they hoped to provide services on the product side as soon as possible, but still hoped to gain more space for self-developed large-scale products.
And Byte’s confidence in facing the problem of accessing DeepSeek sends a positive signal to the outside world.
According to late reports, at the All Hands meeting held on February 13, CEO Liang Rubo believed that Byte did not pay attention to the long-chain thinking model and did not complete the reproduction as soon as possible, which was a decision that needed reflection last year. In 2025, an important goal for the Byte AI team is to “pursue the ‘intelligence’ upper limit.”
In addition to the big factory, the Dark Side of the Moon has also made a similar choice to Byte. The newly released Kimi k1.5 model also has long-term thinking capabilities. Kimi member Flood Sung mentioned in an article reviewing Kimi k1.5 that when the team used RL (Reinforcement Learning) to train LLM (Big Language Model) to do questions through Long-CoT (Long-Chain of Thought), it gave up the PRM model of the reward generation process and instead only rewarded results. “No matter what goes wrong in the model, as long as it is not repeated, if the model is done correctly in the end, we think this is a good exploration and deserves encouragement. On the other hand, if the model is explored and eventually does something wrong, no matter how hard you work, it is still a mistake and you will need to be punished.”
During practice, Kimi’s team discovered a similar phenomenon to the DeepSeek team, that is, under intensive learning, the model will continuously increase the number of tokens for long thinking as training improves performance, and emerge.
Back to the original starting point, how does DeepSeek view large manufacturers joining their own models and “distributing” their own DAUs?
Liang Wenfeng may be happy to see it happen.
From the moment it chose open source, DAU was destined to become a meter of DeepSeek’s model capabilities, not a meter of how much money it could make. The influx of a large number of users will also bring overloaded servers, high costs, and unresponsive “if the server is busy, please try again later”. To this day, DeepSeek’s API service recharge is still suspended.
Access to platforms such as Tencent distracts users ‘access to DeepSeek and also shares the pressure on DeepSeek’s servers. A team that is good at iterating models focuses on the development and iteration of models, allowing people who are good at making entrances to do entrances, people who are good at searching to do searches, and people who are good at making products to do products. It seems that this story is harmonious like an Internet version of a fairy tale.
But LLM will not stop there, and there is more than LLM in the big model. The next moment of emergence will continue to appear in the deep search or the dark side of the moon, in ByteDance or Alibaba, and in Baidu or fast hands. No one can be sure. Standing at a crossroads, no matter what choice the big manufacturers make, it deserves more patience and waiting.
The emergence of an anti-DAU narrator pushed everyone a step forward. At a moment when you have a competitive advantage, switch to the Chinese Internet and first dream of open source.
