Why is the most powerful "grass planting" empire suddenly accelerating its embrace of AI?
Author | Xiao Sijia
Editor | Qiao Qian, Yang Xuan
All major Internet companies aspire to secure a position in the AI era. In this race, Xiaohongshu used to be the cautious one.
In a product that combines search and community attributes, centered around real - life experience sharing, the sense of real people and AI, as well as warmth and algorithms, are always like the two ends of a balance.
For a long time, Xiaohongshu neither completely abstained from technological exploration nor aggressively promoted the productization of AI like many of its peers. Instead, the company has always advanced through the tug - of - war and balance between two forces: continuously investing in model capabilities on one hand, and carefully controlling AI's involvement in the community ecosystem on the other.
However, in 2026, with the rising popularity of the Agent narrative, Xiaohongshu began to show a sense of urgency.
On April 30th, Xiaohongshu sent an internal letter to all employees, announcing the establishment of the first - level AI department Dots, aiming to "build a complete technical system from model R & D, infrastructure, engineering to products, and integrate top - notch AI talents and resources." Dots reports to Conan, the new president of Xiaohongshu.
According to 36Kr, the Dots department was upgraded from the former Humanities Intelligence Laboratory Hi Lab, and it consists of four sub - departments: model R & D, infrastructure, engineering, and products. Currently, "Diandian", the most important AI application product within Xiaohongshu, is also incorporated into this system.
Behind the decision to prioritize AI is a long - standing unsolved problem for Xiaohongshu. Over the years, Xiaohongshu has always held community assets and an advertising base that are envied in the industry. It has also tried to drive new growth through e - commerce, live - streaming and other businesses, but has never really broken through to a new level of business scale.
The "community" is like a treasure chest full of possibilities, and AI is regarded as a new variable and opportunity. However, it remains unknown whether it can be the answer. At least for now, Xiaohongshu's AI story is still in its early stages.
01 An Unsuccessful Productization Path
When ChatGPT was launched at the end of 2022, regarding Xiaohongshu's future, founder Mao Wenchao borrowed an employee's phone and asked GPT a question: Will Xiaohongshu be disrupted? What kind of risks will AI bring to Xiaohongshu?
According to an insider who revealed to 36Kr, during that period, Mao Wenchao began to closely follow AI - related news and required the team to report progress every two weeks.
In an internal letter in August 2023, Mao Wenchao re - emphasized Xiaohongshu's core advantages. He mentioned that when communicating with foreign friends, he found that people would ask a lot of life - experience - related questions on ChatGPT, which overlapped with Xiaohongshu's positioning. However, he also judged that this phenomenon was due to the lack of similar experience content precipitation overseas, while Xiaohongshu had a large amount of real - life content sharing. In other words, the content moat that Xiaohongshu has built up over the years is not easily shaken by AI for the time being.
However, from another perspective, Mao Wenchao did regard AI as a potential threat to Xiaohongshu.
The emergence of ChatGPT quickly impacted traditional search companies like Google and Baidu. For Xiaohongshu, which still had the slogan "Your Life Guide" at that time and emphasized its community search attribute, it also created an unavoidable sense of unease.
This unease led to action. Catalyzed by ChatGPT, in 2023, domestic Internet giants accelerated the layout and investment in large - model bases and gradually promoted productization, resulting in the current AI application forms represented by Doubao, Qianwen, and Yuanbao. The aim was to seize the entrance to the next era.
Xiaohongshu also considered this path. In 2023, Xiaohongshu initially started to promote the construction of technical capabilities internally. In March, it prepared an independent large - model team, led by Zhang Debing (with the Xiaohongshu name "Yuchen"), to self - develop a general large - model base. The team initially focused on base - model R & D and then gradually extended to the product side. In August 2024, it launched an independent AI application, Diandian.
However, different from companies with a certain technical gene, this was a brand - new and unfamiliar path for Xiaohongshu.
Diandian was positioned as a life search assistant and was initially connected to Xiaohongshu's self - developed Zhujji large model. However, on February 26, 2025, Diandian publicly announced that it would connect to DeepSeek. According to the recollection of the aforementioned person, this was because the actual user experience of its basic model was not ideal.
According to another insider, the Diandian team was initially led by Yuchen and reported directly to Mao Wenchao. However, in early 2025, Xiaohongshu re - split the team: the technical side was still led by Yuchen, while the product side was taken over by Tian Sheng, a Stanford alumnus of Mao Wenchao who newly joined Xiaohongshu. However, the Diandian team has undergone several rounds of changes, and currently, Tian Sheng serves as the person in charge of post - training of Xiaohongshu's basic model.
To some extent, this also announced that Xiaohongshu's self - developed model path has not been successful for the time being. "The large - model department has not been abolished, but it has no contact with various products," said the aforementioned insider. A proof is that another AI search product "Wen Y Wen" in the Xiaohongshu community is connected to the open - source Tongyi Qianwen model.
Along with the adjustment of model selection, Diandian also showed a more obvious shift in its product path.
According to 36Kr, Diandian's exploration has roughly gone through four stages:
First stage: Base - model exploration. The early team focused on self - developed basic models, but the results did not meet expectations, and finally they turned to connecting to external models such as DeepSeek. However, the internal exploration of the base - model path has not completely stopped.
Second stage: Humanistic turn. Mao Wenchao proposed to create a more humanistic persona for Diandian, and put forward persona portraits such as "looking up at life at a 45 - degree angle" and "Platonic traveler". Meanwhile, in early 2025, Xiaohongshu integrated the original large - model technology and application product teams into Hi Lab (Humanities Intelligence Laboratory) and recruited a large number of "humanities trainers" with humanities backgrounds for post - training of the model.
Third stage: Product iteration. In July 2025, Diandian launched version 2.0, upgrading the UI and adding multi - modal interactions such as voice and image. However, the version upgrade also caused user disputes. For example, the dark background and bubble - style answer structure were considered to reduce the reading experience.
Fourth stage: Return to the community. On November 28, 2025, Diandian's official account posted two consecutive announcements, declaring the completion of key adjustments: not only restructuring the app's form, but also entering the top - left entrance of Xiaohongshu's main site for the first time, shifting from an independent application to an in - app built - in function and becoming an auxiliary tool for daily browsing and searching.
However, the progress of this product did not meet expectations. "It has not differentiated itself from other AI - based dialogue products on the market and lacks competitiveness," said several people close to Xiaohongshu.
On the user side, Diandian has always had a low profile and hardly made any splash. As of the end of April 2026, in the App Store's download ranking list, Diandian only ranked 186th. In the rating and review section, only 45 users left ratings for it, while the data for "Doubao" was as high as 1.92 million.
Integrating into the community is both a step back and a step forward for Diandian. Since its launch, Diandian has tried several times to integrate with the community ecosystem, increasing its presence in front of users through forms such as "Ask Diandian" or "interaction in the comment section". Xiaohongshu's high - quality community traffic is undoubtedly fertile soil for its development.
However, besides Diandian, there is already another AI search assistant named "Wen Y Wen" in the Xiaohongshu community. There is no obvious difference in their functions, as both generate answers to user questions based on community data. This product was incubated by the search team under Xiaohongshu's community department and reports to Xiahou, the person in charge of algorithms.
The redundant construction at the product level is essentially different departments' separate bets on the AI path. This "horse - racing" mechanism is not uncommon within Internet companies. According to 36Kr, Xiaohongshu is currently considering integrating the two products. And like the integration, the key question of the exact position of AI within Xiaohongshu remains undetermined.
02 Caution and Hesitation in a Real - Person Community
In the face of the new technological wave, the FOMO (fear of missing out) emotion has not been absent within Xiaohongshu. However, compared with the industry's eagerness to embrace AI, Xiaohongshu's attitude towards AI has been extremely cautious - even vigilant for a long time.
In an internal meeting of the community search team in 2024, an algorithm engineer raised a question: Will AI search replace traditional search, or what is the relationship between the two? Xiaohui, the person in charge of community search algorithms at that time, responded: "This is worth trying, but it won't be my top priority."
Looking at the bigger picture, this logic still holds. Compared with AI, Xiaohongshu has more urgent matters to attend to. Before 2026, commercialization and e - commerce were always the top priorities within Xiaohongshu, taking up the top two positions in the OKR, while AI could only rank third. Beyond the priority, there is a deeper concern: for a platform that relies on the community atmosphere and the "sense of real people" to operate, AI and the content it generates lack warmth. Will users naturally reject it? At that time, there was no clear answer to this question.
"The design of the non - aligned double - column pictures on Xiaohongshu's home page is to create a sense of a city street. Users are like strolling around in a neighborhood, seeing unevenly arranged signs. Clicking into a post is like knocking on a door," a former Xiaohongshu community - line employee told 36Kr. "Therefore, Xiaohongshu has always been hesitant about where to place the AI entrance."
This hesitation is also reflected in the change of Diandian's position in the community. Li Shu, a former Xiaohongshu search employee, recalled that when Diandian first entered the community, its entrance was placed in the fourth column of the sug on the search page. "They always wanted to enter the community's home page, but the opportunity never came. At that time, the internal judgment was that there should be no AI in the community."
In the process of promoting another AI product, "Wen Y Wen", Li Shu's team also faced similar choices. When users initiate a search, there are disputes about which questions are suitable for popping up AI summaries and the scope of coverage. Li Shu said that initially, "Wen Y Wen" could only cover 1% - 2% of queries (user search requests). "At that time, we hoped to expand it to 10%, but the boss disagreed, arguing that the impact was too large, and finally it was only relaxed to 3% - 4%."
In 2024, there was no disruptive innovative product in the entire AI market. After the initial anxiety gradually subsided, Mao Wenchao once lost interest in AI, and the frequency of related reports changed from once every two weeks to once a month or even once every two months. "Later, he didn't listen to the reports very much," the aforementioned insider revealed. At that time, Mao Wenchao believed that Xiaohongshu's focus should return to its own business. Although the internal AI projects were still being promoted, they did not receive extra attention.
It wasn't until early 2025 that Deepseek emerged, and China welcomed its own GPT moment. This meant that it became a barrier - free thing for domestic users to use AI.
There was a fierce discussion within Xiaohongshu, and they even conducted a special data comparison: if users installed both Deepseek and Xiaohongshu on their phones, would the search volume of Xiaohongshu decrease? "During the ChatGPT era, they didn't think AI would replace traditional search, but after Deepseek came out, they really started to worry about this," Li Shu recalled.
Regarding "whether to connect to the Deepseek R1 model like Tencent's Yuanbao", the "Wen Y Wen" team had a lot of debates. The back - end technical staff also repeatedly tested the effects after connecting to R1, constantly analyzing the quality of the model's output answers, waiting time, and privacy sensitivity to determine whether to implement this plan.
However, this decision was ultimately not adopted. The reason was still user experience - although connecting to Deepseek could produce higher - quality answers, due to the longer thinking chain of R1 at that time, the time to generate answers increased. Xiaohongshu believed that this would affect users' experience in the community.
Looking back now, Li Shu believes this was a very correct decision: "Speed is the most crucial factor for AI search. Because users won't wait, and they will think it's a bug."
There are also skeptical voices. Some employees think that connecting to Deepseek can obviously increase the richness and sense of real people in the answers. Why not use it? They thought of many ways to try to solve the problem of waiting time, such as offline production - generating answers in advance and directly showing them when users search.
However, these ideas were not actually implemented. Besides the concerns of the decision - making layer, there were other reasons. "At that time, there were many other tasks to be busy with, so no one was available to promote this," Li Shu said. During the Spring Festival in 2025, Xiaohongshu has always had the tradition of CNY (Spring Festival marketing). It not only sponsored the Spring Festival Gala but also planned various activities. The purpose was to spread the brand's youthful and social image and resonate with users emotionally. These tasks occupied most of the human resources and employees' attention, and AI still did not have a high priority at that time.
Caution itself is a signal. In 2022, after nine years of its establishment, Xiaohongshu first opened its community traffic to the e - commerce business. Facing the AI wave, Xiaohongshu has maintained a consistent attitude: it would rather be slow than make mistakes. Therefore, Xiaohongshu has been exploring how to make AI - generated answers more "human - like" and more in line with the community atmosphere.
Besides the user experience, another important question is: in a community product, if an AI search summary tool is implanted, will it reduce users' app usage time? When users can directly obtain answers and guides in a more efficient way, is it still necessary for them to spend time browsing Xiaohongshu?
However, facts have proved that the management's initial concerns about AI destroying the community have not come true. Instead, AI search has brought an increase to the Xiaohongshu community.
According to 36Kr, in 2025, the "Wen Y Wen" function increased the community user retention rate by about 2% - 3% - which is a significant increase for a mature product. Currently, the latest daily active user volume of "Wen Y Wen" reaches tens of millions. The user query volume has also increased by millions. "The browsing time for a single question may have decreased, but the overall usage frequency will increase," Li Shu summarized. "For example, someone who used to ask only one question a week now asks questions every day, so the user retention of Xiaohongshu will increase."
"Data" is one of the core assets of language models. The real content accumulated by Xiaohongshu forms a natural high - quality corpus. On this basis, the generation and summarization capabilities of AI search can enable users to reach answers more quickly. This also makes it a logical choice for "Diandian" to enter the community's home page at the end of 2025.
However, conflicts with commercialization have also emerged. Since AI search generates product recommendations based on real discussions among users in the community, these "native results" are likely to conflict with the interests of brand owners.
For example, when a user asks "What hair dryers are good?", AI search will automatically generate three brands A, B, and C, and the percentage of each brand being recommended in the community. "If a major client has placed a commercial advertisement on Xiaohongshu but is only ranked third in the AI search, the client will definitely be unhappy," Li Shu said. The commercialization team once hoped that the AI search team would not actively generate shopping recommendation lists. However, this function is still displayed on Xiaohongshu.
The contradiction between "realism" and "commercialization" has always been a