Is Xiaohongshu actually in a rush to develop AI?
When Xiaohongshu ventures into AI, the outside world mostly interprets its approach as "restrained".
36Kr mentioned that for a long time, Xiaohongshu has neither been absent from technical exploration nor pushed AI productization in a high-profile manner like its peers. It keeps investing in the construction of model capabilities while carefully controlling the intervention of AI into the community ecosystem.
Many people regard this balancing act as "the unique aesthetic of composure that only slow-growing companies possess".
From my perspective, this is more like a catch-up forced by external forces. It only starts to run when it finds the road ahead is almost blocked.
Two Years of Awkward Contradiction
Xiaohongshu's attitude towards AI in the past two years has been extremely fragmented and full of awkward contradictions.
In 2024, no highly disruptive products emerged in the AI market, and Mao Wenchao's anxiety about AI eased accordingly, with the frequency of internal reports dropping sharply. An insider disclosed to the media that he believed Xiaohongshu's focus should return to its own business. The AI projects are still being promoted, but no extra attention is paid to them.
Compared with AI, Xiaohongshu assigns higher weight to commercialization and e-commerce. One concern speculated by the media is that its operation relies on the community atmosphere and "real human presence", while AI and its generated content lack warmth, which may trigger natural rejection from users.
Against this background, the independent AI application "Diandian" was launched, trying to seize user entrances just like Doubao, Qianwen and Yuanbao. However, the path did not work out, and the effect was far from satisfactory.
When building an independent entrance, it cannot compete with large tech giants; when returning to the main platform, it is just a regular function. The embarrassment of Diandian has previewed the dilemma of Xiaohongshu's AI layout in advance.
The turning point for Xiaohongshu's AI business came in the autumn of 2025. After Mao Wenchao finished his inspection trip in Silicon Valley, he set the tone in a closed-door meeting: the AI technology route has converged, and more investment can be made.
The so-called "convergence of technology routes" means that the industry has gradually formed a general model for AI implementation: based on large models, supporting multi-modal input, equipped with vertical data knowledge base and RAG, and finally forming Agent-style interaction.
This was the stimulus brought by DeepSeek. Xiaohongshu's internal actual measurement found that user search volume was genuinely diverted, so it had to speed up its AI layout.
In the past few years, there are no more than three reasons for Xiaohongshu's awkwardness in its AI business.
First, the top leader did not have a strong enough determination for AI. At the end of 2022, when ChatGPT came out, Mao Wenchao once borrowed an employee's mobile phone to ask GPT a question: Will Xiaohongshu be subverted? During that period, he paid close attention to AI dynamics intensively and required the team to report every two weeks.
However, in the internal letter in August 2023, he reaffirmed Xiaohongshu's core advantages. In his view, people ask life experience questions on ChatGPT because there is a lack of such content precipitation overseas, while Xiaohongshu has a massive amount of real sharing, which he regards as the content moat that cannot be easily shaken by AI for the time being.
In other words, he regarded AI as a potential threat, but not as a battle that must be fought for.
Questions are questions, answers are answers. In the two years after asking that question, Xiaohongshu just carried on its original business as usual.
This kind of swing is directly reflected in the organization: AI had no unified management for a long time, until the architecture adjustment, when Conan integrated the technical system. But this was also interpreted by the outside world as "focusing on product innovation rather than technological breakthrough".
Second, path dependence, which is also a historical burden. On one hand, the community gene makes it dare not act rashly. The entrance of Diandian was long placed in the fourth column of the search page; the query coverage of "Ask a Question" was gradually released from 1%-2% to 3%-4%. This is essentially out of fear that the community atmosphere will be irreversibly damaged.
Xiaohongshu was also hesitant in choosing the basic model. It first developed a self-developed model, then chose to access DeepSeek, but later abandoned it. This instead shows that when there is a conflict between model performance and product experience, Xiaohongshu does not have the ability to bridge the gap through engineering optimization, and can only choose a downgraded solution. "Restraint" in this context is actually limited by capability, meaning "unable to do better".
On the other hand, its corpus assets make it feel no need to rush. The real content precipitated from 250 billion searches on Xiaohongshu is indeed one of the best training corpora of this era.
What large model giants lack is exactly the data with a strong sense of daily life, which is why in the past six months, Doubao and Qianwen have successively launched new voice input methods. They are targeting the most real, unmodified human expression data across the entire network.
But the corpus is both honey and shackles. Whether this asset can operate independently without the community, Diandian has already given an answer: as an independent App, it only ranked 186th on the download list, with only 45 raters, and finally had to return to the main platform.
The AI value of Xiaohongshu is highly dependent on the community context, and it will lose competitiveness once separated from it. The so-called path dependence was confirmed only after hitting the wall once.
Xiaohongshu also has a systematic contradiction. AI search sorts results based on real discussions, which directly conflicts with the advertising placement positions of big clients. The commercialization team once required AI not to actively generate shopping lists.
The contradiction between "real community" and "commercial monetization" has been magnified in the AI era. There is no technical solution, and it can only be resolved through slow interest game and rule reconstruction. This knot cannot be untied by models, but by people and time.
Third, strategic fluke leads to sluggishness in the evolution route of AI. Before 2025, there was a consensus within Xiaohongshu: AI applications like ChatGPT will not replace community search. But DeepSeek shattered this "fantasy".
By 2026, the industry narrative has turned a new page. Chat products with conversation as the core have stepped back, while OpenClaw ("Lobster") has pushed Agent (intelligent agent) to the forefront and become the new center of attention. The FOMO sentiment of the entire Internet industry has risen sharply, and even the most conservative organizations cannot deny the importance of AI.
It was at this point that Xiaohongshu realized it was half a beat slow again. If it does not catch up quickly, it may not even be able to touch the poker table.
With the three factors superimposed, it cannot speed up even if it wants to, unless the top management makes up its mind to go all in on AI.
Xiaohongshu's AI Business Has to Run Faster
But now, it has to run faster, and three forces are pushing Xiaohongshu to accelerate passively.
The first force is the real panic caused by the diversion of search entrances. Search is the lifeblood of Xiaohongshu, and 73% of its monthly active users have search behaviors. Xiaohongshu has kept taking actions to defend the search entrance with AI, including the AI chat tool Da Vinci, the Diandian App, and the in-platform "Ask a Question" function, but the effect is minimal. It has always maintained a posture of restrained launch and small-scale coverage, and dare not widely replace the traditional note stream.
The second force is that Xiaohongshu needs an AI story for its listing. In the second half of 2025, there were once rumors that Xiaohongshu would go public in April this year. A former employee said, "Since it has not gone public, Xiaohongshu needs the AI story even more."
AI plays a dual role: as a technical tool on the surface, and as valuation support in essence. When business growth cannot meet expectations, AI has to step up to fill the gap. This eagerness is so full of capital flavor that it cannot be hidden.
The third force is the "performative" suspicion of organizational mobilization. In March 2026, the internal slogan of "all staff transform into AI product managers" was put forward. Non-technical positions were required to learn vibe coding, and only AI positions were opened for campus recruitment with top salaries.
It should be noted that large tech giants invest hundreds of billions of money in basic models to bet on the next era, while what Xiaohongshu launched is just an organizational mobilization. This contrast itself explains the problem: what it fears is not spending money, but being excluded from the game.
Now, Xiaohongshu's passive responsive operation mode has split into three routes.
The clear route is to reinforce the community moat with AI. The daily active users of "Ask a Question" have reached the order of ten million, with retention rate increased by 2%-3%, but the answers must come from real notes. The logic is always "embedding" rather than "replacement".
The hidden route is just to ensure its presence: the open-source model Dots has received mediocre feedback in the industry, and the confidential project team deliberately hides Xiaohongshu's identity. The self-developed model is the admission ticket, and the confidential project is the error-tolerant probe, which guarantees its qualification to participate in the next round of technological change.
The middle zone is horse racing: Diandian and "Ask a Question" have overlapping functions, and the integration plan is still pending. It is like a company without a commander, where each platoon digs its own trenches.
The clear route maintains the existing advantages, the hidden route explores new paths, and the horse racing mode fills the gaps. Companies that have not figured out their strategy will act in a scattered way when they are anxious. The three routes have their own separate accounts, but there is no general account at all.
Besides, Xiaohongshu's shackles are still on it. For example, in terms of computing power, its internal statement is that "there is no shortage of GPUs for simple tasks, but it may be a little insufficient if we want to carry out large-scale operations".
This "little insufficiency" directly determines that Xiaohongshu cannot join the arms race of basic models, and can only find differentiation in the application layer and vertical scenarios.
The talent threshold is also difficult to cross. The CTO position of Xiaohongshu has been vacant since 2020. Its engineers are described as "one person can do the work of three", which shows good efficiency, but AI competition relies on talent density and technical depth.
Another point is that in terms of governance, community security always comes before AI expansion. For example, more than 300,000 pieces of content are labeled every day, and nearly 180,000 pieces of AI-generated fake content have been disposed of in total, but the boundary between "AI-assisted creation" and "AI-generated content" still does not have a clear standard to this day.
Therefore, when Xiaohongshu develops AI, it is not that it does not want to speed up, let alone being "restrained". The awkward performance it shows is because it can only move forward with brakes on.
This is a sober AI survival strategy, and also a pragmatic choice made by a company with limited resources, special genes and pushed by external pressure under multiple constraints.
Whether Xiaohongshu can avoid being eliminated in the AI race depends on two variables: Will AI-generated content erode the scarcity of "authenticity"? Can the barriers of vertical scenarios resist the dimensionality reduction strike from tech giants?
But one thing is clear: Xiaohongshu, the company that once believed in "slowness", has been pushed onto the fast track in the AI era. Once it gets on the bus, there is no option to get off.
It still needs to find the answer to this question by itself: when the community gene meets the AI wave, is restraint a amulet, or a ceiling?
References:
36Kr, "Xiaohongshu's Four-Year AI Journey: FOMO, Hesitation, to Sudden Acceleration"
Tang Chen Classmate, "Is the Input Method Game Table Going to Be Overturned by Doubao and Qianwen?"
This article is from the WeChat official account "Tang Chen Classmate", author: Tang Chen, published by 36Kr with authorization.