Four years after ChatGPT, Xiaohongshu’s AI restraint gives way to urgency

Internet companies are finding it hard to resist the appeal of the artificial intelligence boom. But for some, AI sits at the opposite end of their product logic, requiring restraint to avoid disrupting the foundations of their business.

Xiaohongshu, which operates internationally as RedNote, falls into this category. The platform combines search with community features and is centered on the sharing of real-life experiences. For a product built around human warmth, presence, and community, AI can appear to be in tension with the atmosphere Xiaohongshu seeks to create online.

But Xiaohongshu has not shied away from turning AI into products. Instead, it has sought to balance the two forces, continuing to invest in AI model capabilities while carefully controlling the technology’s role in its community ecosystem.

In 2026, as the agentic AI narrative gained momentum, Xiaohongshu began to show more urgency.

On April 30, Xiaohongshu sent an internal letter to all employees announcing the establishment of Dots, a first-level AI department. The department is intended to build a technology system spanning four divisions: model R&D, infrastructure, engineering, and product. It will also integrate top AI talent and resources. Dots reports to Xiaohongshu’s newly appointed president, Ding Ling, also known by the alias Conan.

According to 36Kr, Dots was upgraded from Hi Lab, formerly the Humane Intelligence Lab. Xiaohongshu’s key internal AI application, Diandian, has also been placed under this system.

Behind the decision to elevate AI lies a long-running question Xiaohongshu has yet to solve. For years, Xiaohongshu has had community assets and a core advertising business that peers envy. It has also tried to unlock new growth through e-commerce, live streaming, and other businesses. But it has never fully broken through or taken its business scale to another level.

Why Xiaohongshu’s productization strategy has yet to work

When ChatGPT emerged in late 2022, Xiaohongshu founder Mao Wenchao, also known by the alias Seiya, once borrowed an employee’s phone and asked the chatbot about Xiaohongshu’s future. Would Xiaohongshu be disrupted? What risks would AI bring to the platform?

People familiar with the matter told 36Kr that, during that period, Mao began paying close attention to AI-related developments and required the team to provide progress updates every two weeks.

In an internal letter in August 2023, Mao reiterated Xiaohongshu’s key strengths. He said that while speaking with foreign friends, he found that people were asking ChatGPT many questions related to life experiences, which overlapped with Xiaohongshu’s positioning. But he also judged that this phenomenon stemmed from the lack of a similar accumulation of experience-based content overseas, while Xiaohongshu had a vast amount of content rooted in real-life sharing. The implication was that Xiaohongshu’s content moat would not be easily shaken by AI in the short term.

From another perspective, however, Mao did see AI as a potential threat.

The emergence of ChatGPT quickly unsettled traditional search companies such as Google and Baidu. For Xiaohongshu, which at the time still emphasized community search, this created a sense of unease that was difficult to ignore.

That unease led to action. Catalyzed by ChatGPT, China’s internet giants accelerated their deployment of and investment in foundation models in 2023. They then gradually moved toward productization, leading to app formats now represented by ByteDance’s Doubao, Alibaba’s Qwen, and Tencent’s Yuanbao. The goal was to seize the entry point to what many saw as the next era of consumer internet.

Xiaohongshu also considered this path. In 2023, the company began building internal technical capabilities. In March, it prepared an independent large model team led by AI innovation head Zhang Debing to develop a general-purpose foundation model in-house. The team initially focused on foundation model R&D, then gradually extended toward products. In August 2024, it launched Diandian as an independent AI application.

But unlike companies with deeper technological backgrounds, this was a new and unfamiliar path for Xiaohongshu.

Diandian was positioned as a lifestyle search assistant. At first, it was connected to Xiaohongshu’s self-developed Zhuji large model. But on February 26, 2025, Diandian publicly announced that it had connected to DeepSeek. According to the people cited above, this was because the actual user experience of its foundation model was found to be unideal.

Another person familiar with the matter said the early Diandian team was led by Zhang and reported directly to Mao. But in early 2025, Xiaohongshu split the team again: Zhang continued to oversee technology, while the product side was handed to Mao’s Stanford classmate and a new Xiaohongshu hire who goes by the alias Tiansheng. The Diandian team later went through several rounds of changes. Tiansheng now serves as head of post-training for Xiaohongshu’s foundation models.

To some extent, this also signaled that Xiaohongshu’s self-developed model path had not yet worked. “The large model department has not been abolished, but it also has no contact with the various products,” the person familiar with the matter said. One piece of supporting evidence is that another AI-driven search product inside Xiaohongshu’s community, Wenyiwen, uses the open-source Qwen model.

As Xiaohongshu adjusted its model choices, Diandian also made a more visible shift in product direction.

According to 36Kr, Diandian’s exploration has gone through roughly four stages:

But the product’s progress has fallen short of expectations. “It has not differentiated itself from other AI chat products on the market and lacks competitiveness,” several people close to Xiaohongshu said of Diandian.

On the user side, Diandian has never had a strong presence and has barely made a splash. Based on an end-of-April review by 36Kr, Diandian ranked only 186th on the app store download chart. In the ratings and reviews section, only 45 users had rated it, while Doubao had as many as 1.92 million ratings.

Integrating into the community was both a step back and a step forward for Diandian. Since its launch, Diandian has tried several times to connect with the community ecosystem through formats such as “Ask Diandian” and interactions in the comments section, hoping to increase its visibility to users. Xiaohongshu’s community traffic is undoubtedly fertile ground for it.

The issue is that, aside from Diandian, another AI search assistant called Wenyiwen already exists within Xiaohongshu’s community. The two products have no obvious functional differences. Both generate answers to user questions based on community data. Wenyiwen was incubated by the search team under Xiaohongshu’s community department and reports to Xiahou, the alias used by the head of algorithms.

Duplicate product development is, at its core, the result of different departments placing their own AI bets. This kind of practice is not unusual inside internet companies. 36Kr has learned that Xiaohongshu is currently considering integrating the two products. But another key question, as unsettled as the integration itself, is what position AI should ultimately occupy inside Xiaohongshu.

Restraint and hesitation within Xiaohongshu

Faced with a new technology wave, Xiaohongshu was not immune to the fear of missing out. But compared with the bandwagon effect AI has had on many other companies, Xiaohongshu’s attitude toward AI remained unusually restrained, even wary, for quite some time.

At an internal meeting of the community search team in 2024, an algorithm developer asked whether AI-driven search would replace traditional search and how the relationship between the two should be understood. Xiaohui, the alias used by the head of community search algorithms, responded at the time: “This is worth trying, but it will not be my O1 or O2,” implying those items were not at the top of the agenda.

The same logic held at a broader level. Compared with AI, Xiaohongshu always seemed to have other priorities it considered more urgent. Before 2026, monetization and e-commerce had long been the company’s main priorities, occupying the top two positions in its objectives. AI ranked third at best.

Beyond priority, the deeper concern was this: for a platform that depends on community atmosphere and a sense of real human presence, would users naturally reject AI and AI-generated content? At the time, there was no clear answer.

“The misaligned two-column image-and-text layout on Xiaohongshu’s homepage is designed to create the feeling of a city street. Users are like people wandering through a neighborhood, seeing signs of different sizes staggered along the way. When they click into a post, it feels like knocking on a door,” a former employee from Xiaohongshu’s community line told 36Kr. “So Xiaohongshu has always hesitated over where exactly to place AI’s entry point.”

This hesitation was also reflected in changes to Diandian’s position inside the community. A former employee from Xiaohongshu’s search department who wanted to be known only by the pseudonym Lishu recalled that when Diandian first entered the community, its entry point was placed in the fourth slot of the search suggestion page. “They always wanted to integrate it into the community homepage, but that opening was never made. The internal judgment at the time was that AI should not appear in the community.”

Lishu’s team faced similar tradeoffs while advancing another AI product, Wenyiwen. When users initiated searches, there were debates over which questions were suitable for AI summaries and how broadly they should be covered. Lishu said Wenyiwen could initially cover only 1–2% of queries, meaning user search requests. “At the time, we hoped to expand that to 10%, but the boss did not agree, saying the impact would be too large. In the end, it was only relaxed to 3–4%.”

In 2024, the overall AI market did not produce any truly disruptive innovative products. After the initial anxiety faded, Mao at one point lost interest in AI. Related briefings changed from once every two weeks to once a month, or even once every two months. “Later on, he slowly stopped listening much,” the person familiar with the matter said. At the time, Mao believed Xiaohongshu’s focus should return to its own business. Internal AI projects were still moving forward, but they did not receive extra attention.

That changed in early 2025, when DeepSeek suddenly emerged and China had its own “ChatGPT moment.” This meant that using AI had become largely frictionless for domestic users.

Xiaohongshu held intense internal discussions and even conducted dedicated data comparisons: If users had both DeepSeek and Xiaohongshu installed on their phones, would their search volume on Xiaohongshu decline? “They did not think AI would replace traditional search. But after DeepSeek came out, they started to truly worry about it,” Lishu said.

The Wenyiwen team debated extensively whether it should connect to DeepSeek’s R1 model, as Yuanbao had done. Backend technical staff repeatedly tested the results of connecting to R1, continuously analyzing the quality of AI-generated answers, waiting time, and privacy sensitivity to determine whether the plan should be implemented.

But the decision was ultimately not adopted. The reason was still user experience. Although connecting to DeepSeek could produce higher-quality answers, R1’s reasoning process was longer at the time, which increased answer generation time. Xiaohongshu believed this would affect the user experience inside the community.

Looking back, Lishu believes this was the right decision:

There were also dissenting voices. Some employees believed that connecting to DeepSeek would clearly increase the richness of answers and the sense of real human presence. Why not use it? They considered ways to solve the waiting-time problem, such as offline generation, which would mean generating answers in advance so they could appear directly when users searched.

But these ideas were never fully implemented. In addition to management’s concerns, there were other reasons. “At the time, there were too many other tasks, so no one was available to push this forward,” Lishu said.

It was around the 2025 Lunar New Year, and Xiaohongshu has long had a tradition of festival-period marketing. It not only sponsors the gala televised by CCTV, but also plans various activities. The goal is to communicate a younger and more social brand image, and to create emotional resonance with users. These tasks occupied most of the company’s human resources and employee attention. AI still did not have priority at that point.

Caution itself is a signal. In 2022, nine years after its founding, Xiaohongshu opened community traffic to its e-commerce business for the first time. Faced with the AI frenzy, Xiaohongshu maintained its consistent attitude: better to move slowly than make mistakes. As a result, Xiaohongshu has been exploring how to make AI-generated answers feel more human and better blend into the community atmosphere.

Beyond user experience, another important question was whether embedding an AI search summary tool into a community product would reduce users’ time spent in the app. If users could obtain answers and guides more efficiently and directly, would they still need to spend time browsing Xiaohongshu?

But the data showed that management’s initial fears about AI damaging the community did not materialize. AI search instead brought incremental growth to Xiaohongshu’s community.

According to 36Kr, in 2025, the Wenyiwen feature increased community user retention by about 2–3%, a significant gain for a mature product. Wenyiwen’s latest daily active user count has reached the tens of millions. User query volume has also increased by several million. “The browsing time for a single question may become shorter, but the overall frequency of use will increase,” Lishu said. “For example, someone who originally asked only one question a week might now ask questions every day. Then Xiaohongshu’s user retention will increase.”

Data is one of the primary assets of language models. The real-life content accumulated by Xiaohongshu forms a natural source of high-quality training material. On that basis, the generation and summarization capabilities of AI search allow users to reach answers faster. This also made it logical for Diandian to enter the community homepage at the end of 2025.

However, conflict with monetization also began to emerge. Because AI search generates product recommendations based on real user discussions within the community, these native results can easily clash with brand interests.

For example, when a user asks, “Which hair dryer is good?” AI search will automatically generate three brands, A, B, and C, along with the percentage of recommendations each brand receives within the community. “If a major client has placed monetized ads on Xiaohongshu but is only ranked third in AI search, the client definitely will not be happy,” Lishu said. The monetization team once hoped the AI search team would not actively generate shopping recommendation lists. For now, this feature is still displayed inside Xiaohongshu.

The tension between authenticity and monetization has always been a recurring problem in Xiaohongshu’s development. Excessive commercialization erodes the community, while insufficient monetization is difficult to sustain. With the arrival of AI, this tension may be further intensified.

Is Xiaohongshu pursuing AI out of fear or strategy?

In the fall of 2025, after returning from a visit to Silicon Valley, Mao said in a closed-door meeting that AI’s roadmap had converged, and Xiaohongshu could invest more.

This so-called convergence refers to the industry’s gradual formation of a common model for AI implementation: a large model as the base, support for multimodal input such as voice and images, a vertical-domain knowledge base and retrieval-augmented generation, or RAG, and finally agent-style interaction. “In other words, this means search is very important, and users may also use multimodal methods such as voice and images to search,” one person familiar with the matter said.

Search is precisely the user association Xiaohongshu has built most strongly beyond its identity as a recommendation-driven community. After a year of exploration around AI search, Mao seems to have gradually realized that AI would not damage the value of the community and that it had use cases that could combine with Xiaohongshu’s search mindshare. As a result, Xiaohongshu’s attitude toward AI shifted from initial wariness and restraint to a more active embrace.

With the arrival of 2026, the industry narrative around AI also turned a new page. Chat products centered on conversation receded into the background. OpenClaw’s emergence pushed a category of agents built on coding capabilities and equipped with task execution abilities to the forefront, making them a new center of attention.

Even conservative organizations can hardly deny AI’s current importance. Many internet companies are trying to prove that they can master the new technology and stay relevant.

Xiaohongshu became anxious as well. “Everyone is transforming into an AI product manager,” one Xiaohongshu employee said.

Since March, employees have been able to feel AI’s position inside Xiaohongshu rising quickly. Technical teams have frequently organized sharing sessions to teach employees how to use AI skills. Non-AI roles have also been pulled in. Some employees have been required to learn how to use vibe coding to build AI tools and embed them into workflows. They even have to evaluate their direct managers’ AI usage capabilities through questionnaires.

Recruitment has also begun to tilt toward AI. For its 2026 campus recruitment, Xiaohongshu has opened almost only AI-related positions, with engineers making up the overwhelming majority. Before that, the total number of employees working on AI internally was no more than around 100. By contrast, ByteDance’s Doubao team alone has more than 1,000 people.

It is understood that when Xiaohongshu recruits AI talent externally, it usually requires candidates to meet three criteria at the same time: young, high-potential, and high-performing at a major tech company. Correspondingly, Xiaohongshu has reportedly been generous with compensation, according to 36Kr.

The company pursues a high employee productivity ratio and a flat reporting structure. An engineer is expected to do the work of three people, and product managers have a broad scope of authority and responsibility. “One person may have to do the work of an entire team.” The thinking behind this is that bloated organizational structures add redundancy and disagreement. “A group of mediocre product managers will only create a complicated product,” the employee quoted above said.

In the second half of 2025, there were rumors that Xiaohongshu would list in April this year. The other shoe did not drop. “Because it did not go public, Xiaohongshu needs an AI story even more,” a former Xiaohongshu AI employee said.

Internal exploration of the model base is still continuing.

In the second half of 2025, Xiaohongshu’s Hi Lab team successively released a series of self-developed open-source large models, including a text large model, a vision-language model, and a document parsing model. It also released open-source models and toolchains covering multimodal reasoning, image editing, deep search, and agents. But like Diandian, the Dots series models did not attract much attention from the industry or technical community.

As a company of relatively limited scale, Xiaohongshu’s weaknesses are clear. It does not have an advantage in technical accumulation or computing resources. Although many internal employees often write about ample GPU supply when posting recruitment messages on Xiaohongshu, the company inevitably looks stretched compared with leading tech giants that routinely invest hundreds of billions of RMB. “For simpler tasks, there is no shortage of GPUs. But if you want to build a foundation model, there may be some gaps,” an internal technical staff member told 36Kr.

Xiaohongshu has not been stingy with recruitment, but it has long lacked a technical leader like ByteDance’s Wu Yonghui or OpenAI’s Yao Shunyu. Since former CTO Xi Xiaohu, whose Xiaohongshu alias was Shanqiu, left around 2020, Xiaohongshu has not appointed another CTO. Its technology architecture is mainly overseen by two technology vice presidents who go by the aliases Fengdi and Kaiqi.

From the outside, this looks like an investment with unclear cost-effectiveness. Apart from a few leading companies, it seems difficult for most players that continue to bet heavily on model bases to obtain equivalent returns.

But inside Xiaohongshu, this looks more like a required question that must be answered. “Without this capability, you cannot get a seat at the table,” several technical staff members said. A person close to Xiaohongshu said the significance of the company’s self-developed models is not necessarily how good they are, but that they allow technology to drive the business and preserve the company’s eligibility to participate in the next round of competition. Without underlying capabilities, it cannot truly understand the boundaries of the technology or make judgments at critical moments.

So far, the iteration of AI products has still been driven more by technology than by product strategy. Every leap in large model capabilities spills over beyond its original boundaries and creates new possibilities. New product opportunities may be hidden in those spillover points.

For example, after Claude Sonnet 3.5’s code generation capabilities improved significantly in 2024, they spilled over into agent capabilities for autonomous task execution. The Cursor team was among the first to seize the opportunity. It identified market demand and optimized its product accordingly, quickly winning recognition in the developer market. “Whoever can tap AI’s potential the fastest and discover its most powerful use cases will succeed. This requires people to have a strong perception of both models and products,” the person close to Xiaohongshu said.

Therefore, even if short-term returns are unclear, this remains an admission ticket Xiaohongshu needs to buy.

On April 30, Xiaohongshu sent an internal letter to all employees announcing a new round of organizational upgrades. In addition to establishing Dots, it set up another department focused on enterprise intelligence, highlighting its expanded AI investment from both product technology and organizational perspectives. It also appointed Conan as president. She will fully integrate the company’s three core businesses, community, e-commerce, and monetization, as well as its technology system, and report to Mao.

Conan graduated from Nanyang Technological University in Singapore with a bachelor’s degree in computer engineering, then entered Stanford Graduate School of Business in 2011 to pursue an MBA. Before joining Xiaohongshu, she worked at Boston Consulting Group, Citi, and other institutions, mainly in consulting and finance. Her career path is similar to Mao’s, who also entered Stanford Graduate School of Business in 2011. Xiaohongshu’s number three employee, Deng Chao, also known by the alias Yingmu, is currently chief product officer and head of the community business. He came from an architecture background.

Judging by management backgrounds, Xiaohongshu’s core decision-making team is not known for technical strength. Conan’s appointment to integrate the company’s technology system is also seen externally as a sign that Xiaohongshu’s AI strategy leans more toward product-level innovation than technological breakthroughs.

According to 36Kr, in addition to the AI assistant embedded in the community and Dots, which focuses on underlying R&D, Xiaohongshu also has other undisclosed project groups exploring AI products. These groups are confidential, do not appear in the public organizational structure, and are not aware of one another. Some projects have already incubated products, but when promoted externally, they have deliberately obscured Xiaohongshu’s identity.

For now, Xiaohongshu seems to be pragmatically focused on specific scenarios, such as creating products in verticals like travel and fashion that can solve user needs and capture mindshare in niche segments. It also has assets that are difficult to replicate: a highly active community and users close to real life. This is where lifestyle trends that swept through young people, such as camping and frisbee, once took root.

Compared with chasing technology for its own sake, these assets may give Xiaohongshu a chance to find differentiated opportunities in the next round of AI competition. Yet beyond its existing community narrative, Xiaohongshu eventually needs a new story to prove itself in the AI era.

KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Sijia for 36Kr.

...