After departing from Moonshot AI, he leverages AI technology to help people find their ideal romantic partners, and has secured investment from Xu Xin | Newly Emerged Project

Text by WEN Lihong

Edited by ZHANG Yuxin

On the eve of the launch of Kimi K3, Zeng Xunxun did the math: the value of the stock options he gave up when he left his previous job had risen 10 times during his absence.

Before departing from Moonshot AI, Zeng Xunxun was the technical lead of Kimi's AI search team. In August 2025, he left to found "MatchPerfect Technology", whose core product "MatchPerfect" is an AI-driven dating matching application. Its core logic is to integrate AI search and matching capabilities into the process of "finding people", using AI to boost efficiency in three key stages of the dating scenario: profile collection, intelligent matching, and communication assistance.

Shortly after the company was founded, Zeng Xunxun secured a $2 million angel investment from Xu Xin of Capital Today. He later learned that Xu Xin had vetted the "AI person-finding" direction through several rounds of screening before meeting him. At that time, Xu Xin was focusing on two tracks: "AI job hunting" and "AI partner finding".

In the recruitment sector, she had spoken to almost all notable domestic startup teams in the field, but did not invest in any of them. Xu Xin explained that none of these teams demonstrated a deeper understanding of the recruitment process than previous generations of product managers. No team could provide her with an unexpected answer about which links AI could generate incremental value and how it could restructure matching efficiency.

For the dating sector, Xu Xin only met with the MatchPerfect team. Before investing, Xu Xin mentioned that she hoped to find a founding team whose understanding of "AI-powered person search and matching" exceeded her own insights and that of the companies she had previously invested in. If the team's understanding was merely on par with hers, there would be no value in placing a bet.

Currently, the MatchPerfect mini-program is online, and the full App is about to launch.

Zeng Xunxun, founder of MatchPerfect, speaking at the 2026 World Artificial Intelligence Conference (WAIC). Source: WAIC official

A Startup Built on Familiar Technical Foundations

Zeng Xunxun believes he is entering a market with proven real demand. He cited data from several similar products as reference: the first platform sets its annual membership fee at 1288 yuan, while on the second platform, each paying user spends approximately 800 yuan over their entire lifecycle on the service.

An annual fee of 1288 yuan is at the top end of pricing for consumer subscription products. Compared with video platforms that charge two to three hundred yuan per year, and music platforms that charge just over one hundred yuan per year, the fact that users are willing to pay 1288 yuan for a dating product shows that "finding a partner" ranks extremely high in users' value priorities.

At the same time, past users' common concerns about dating platforms are that they pay high fees but still struggle to accurately and efficiently find the right match. On this point, Zeng Xunxun's judgment is: The technical capabilities of the previous generation of dating products cannot deliver the "precise matching" service. Platforms built huge user pools to gather people with dating needs, but users still had to spend massive amounts of time and effort on the process of finding a partner. Users had to browse, analyze profiles, communicate, match, and filter candidates on their own, repeating the cycle over and over. The recommendation systems barely helped in the task of matching people more accurately.

The core logic of "MatchPerfect" is precisely solving the "precise matching" problem, which is highly aligned with Zeng Xunxun's previous work at Kimi.

While serving as the AI search technical lead at Kimi, Zeng Xunxun's daily work involved leading his team to explore how AI could help users accurately match information, and discovering the commercial value that could be created from this simple capability.

Over time, he observed a phenomenon in AI search: in the general search field, as technology improves, search results get better and answer quality rises significantly, but the computing power costs also rise sharply. However, users are rarely willing to pay for a single search. "Right now, everyone's AI search is provided for free, users can use it without paying token fees, because people don't think this service is worth spending money on, they're reluctant to pay," Zeng Xunxun says.

But specialized AI search products that focus on specific verticals (such as healthcare, finance, law, etc.) are a very different story. These search services often integrate large volumes of private databases to help users search and match information, and users are willing to pay for these targeted results.

Zeng Xunxun argues that this is not only due to differences in technical performance and costs between the two types of search, but also because the answers in vertical scenarios are far more personally relevant: a professional medical answer could save a life, and a professional legal advice answer can directly impact critical decisions.

The high premium in these scenarios perfectly illustrates that it's not that users are unwilling to pay for "information", but that information in general scenarios has too little value. The same AI technology that cannot find a profitable path in general search can find commercial monetization opportunities in specific vertical search scenarios like law and healthcare.

Following this logic, he concluded that "finding people" is another high-value scenario for search. Two typical use cases stand out: recruitment, and dating.

In Zeng Xunxun's view, for AI, finding people is essentially the same thing as high-precision information matching. The technical workflow of AI search and dating matching are completely identical: "understand the intent — analyze and match requirements — output search results". Kimi takes a user's question and matches it with information, while MatchPerfect takes a user's personal profile and their expectations for a partner, and matches them with another person. The target of the search has changed, but the underlying logic remains the same.

By integrating AI search into dating matching, he breaks down the dating matching process into three stages, each addressing a specific bottleneck that has long plagued dating platforms.

The first stage is profile collection. On dating platforms, it's common for users to write self-introductions. But not everyone can describe themselves clearly, logically, and comprehensively, so a large number of user profiles on platforms are nearly blank, leaving the system without sufficient valid information to make quality recommendations.

In Zeng Xunxun's opinion, this is not a problem of user willingness, but a problem of user capability. On "MatchPerfect", the design team built an "AI Matchmaker" to help users complete profile collection. After registering, users have a 20-minute voice call with the AI Matchmaker, which asks personalized questions. Users don't need to carefully organize their language, they just need to answer truthfully. After the call, the AI generates a well-structured self-introduction, sorting out basic information required for dating matching, including the user's relationship expectations and values, to support subsequent demand analysis and matching.

Interface display of MatchPerfect's "AI Matchmaker", image provided by the interviewee

The introduction of AI technology lowers the barrier for users to create high-quality profiles from the requirement of "being able to express themselves precisely". After the MatchPerfect mini-program launched, Zeng Xunxun and his colleagues found that the average length of user profiles was 463 words, while their manual survey of top dating products previously found that the average user profile length was only 132 words.

The second stage is the precise matching algorithm. Previous generation products relied on structured tags like age, height, and education to make recommendations. But in reality, the core factors that determine whether two people are compatible lie in their values, life experiences, personalities, and lifestyles.

All of this information exists in the text users write on their profiles, but traditional recommendation systems could not analyze or leverage this unstructured data. Even after being recommended based on tagged information, users still had to manually filter out a huge number of candidates with mismatched marriage expectations from the crowd, making finding the right person an exhausting process.

Zeng Xunxun is well aware of this: previous generation recommendation algorithms could not process natural language, until the advent of large language models. MatchPerfect is currently training a dedicated matching model, whose input is the user's complete profile, including publicly displayed text and the raw conversation records from their AI Matchmaker calls. After the model understands the content of two users' profiles, it can analyze their in-depth personal information to deliver more accurate matching results.

The third stage is communication mediation. In dating scenarios, there are certain questions that need to be clarified before a relationship gets serious, such as family assets, debts, romantic history, health conditions, and expectations for betrothal gifts. In the early stages when two people first meet, these questions are too awkward to bring up directly. If these issues are delayed for months until after they've invested time in the relationship, the sunk costs will be so high that some people will be unwilling to face the mismatches.

To address this, MatchPerfect has developed two features: AI Avatar and AI Advisor. The first feature allows users to anonymously ask and answer sensitive questions: the recipient of the question does not know who the sender is, they only know that a potential match wants to know specific information about them, which eases the awkwardness of verifying personal details one-on-one. The second feature, AI Advisor, provides conversation topic suggestions and guidance on how to respond appropriately during real-person interactions.

Making a Trade-off Between User Scale and Community Atmosphere

Zeng Xunxun has observed that many dating platforms are forced to waver repeatedly between "user volume metrics" and "community atmosphere".

To expand user numbers, platforms find it difficult to position themselves strictly as "serious dating platforms", and instead generalize the types of relationships they promote, using slogans like "help you make new friends" and "help you find a date". This blurs the platform's boundaries, allowing users with all kinds of different intentions to join. As a result, users with conflicting goals are mixed together on the platform, and ultimately no one is satisfied.

Zeng Xunxun shared a thought with "Intelligent Emergence" that sums up his fundamental judgment on this issue: "If you believe you are building a serious dating app, you should maintain that atmosphere in every single aspect of the product. None of the past products did this well, they all tried to pursue multiple conflicting goals at once, and ended up achieving none of them properly."

His choice was to abandon the "trying to have it all" approach, and clearly define MatchPerfect as a serious dating platform focused on marriage. Aligned with this positioning, he designed screening mechanisms in every stage of the user journey, to make non-target users voluntarily leave during the registration process.

MatchPerfect has implemented multi-layer screening mechanisms: real-name authentication, facial recognition, a roughly 20-minute voice call with the AI Matchmaker, and an AI profile evaluation (users must score above 70 to go online). Users must pass all these steps sequentially to access the platform. The platform also verifies the marital status of registered users, directly blocking married people from joining. In the first week after the product launched, more than 50 married users were blocked from registering.

In addition, MatchPerfect has designed a strict "1v1 matching mechanism". After a user enters a match, they can only chat with one matched person at a time. They can only unmatch and get to know other people after confirming the current match is not suitable. "This is the most important product trade-off we made to maintain a serious community atmosphere," Zeng Xunxun says. "This mechanism can eliminate the suspicion and game-playing that exist in interactions on general social platforms, and ensure that every match is treated seriously by users."

The design of these various rules is based on a logic Zeng Xunxun firmly believes in: "If a person is only looking for a short-term relationship, they will not be willing to reveal so much real personal information on a platform, nor will they be willing to go through so many screening steps."

The community atmosphere in turn constrains user behavior. Zeng Xunxun has verified a judgment through this process: a good community atmosphere is not enforced by rules, but cultivated through screening mechanisms. When the entry threshold is strict enough, after users who meet the standards join the community, a positive atmosphere will naturally form; and once that atmosphere is established, it will in turn guide the behavior of new users. The two elements form a positive feedback loop.

In Zeng Xunxun's view, AI can improve matching efficiency by precisely analyzing users' dating needs. But whether a serious dating community can operate successfully still depends on whether the screening mechanism is strict enough, and whether the business model is aligned with the product's goals. These are things that AI cannot solve; they require human judgment and deliberate implementation. A team that only understands technology will find it difficult to achieve this.

Guaranteed Marriage: A New Attempt at Outcome-Based Delivery

In Zeng Xunxun's opinion, the fundamental reason why previous generation dating products failed was not just backward technology, but a fundamental conflict between their business model and product goals: in the past, almost all dating products used a subscription model, charging membership fees monthly or annually. This means that the longer users stay on the platform, the more revenue the platform generates. As a result, platforms have no real incentive to help users find a suitable partner as quickly as possible.

"In this scenario, platforms are not genuinely motivated to help users succeed. They design all kinds of ways to make users pay, and after extracting the last bit of value from users, they leave them disappointed and ready to churn," Zeng Xunxun says.

Thus, he redesigned the delivery logic of the dating platform. When users pay, what they get is not a time-limited right to use platform services, but a delivery goal centered on "achieving a romantic relationship leading to marriage". Specifically, he planned a "Marriage Guarantee Membership" program, whose core rule is: After paying a 2000-yuan membership fee, if the user does not get a marriage certificate within 3 years of joining MatchPerfect, the platform will refund the full fee.

This model went through three iterations of development.

The first version was "guaranteeing that users meet at least one person". After users sent messages to each other and chatted for a while, arranging an in-person meeting would count as a completed delivery outcome. If the user didn't meet anyone at all, they wouldn't be charged a penny. He discussed this plan with investors, and the feedback was that users might dispute the results: maybe they exchanged WeChat contacts but never met, or they met but the experience was very unpleasant, leading users to refuse to acknowledge that the service was delivered, which made the model too risky.

The second version was "guaranteeing that users can enter a romantic relationship". This had even bigger problems, because being in a relationship cannot be easily verified. Anyone could claim they never dated someone even if they did, making the non-payment rate nearly 100%.

Eventually he realized that marital status is the only verifiable, non-falsifiable delivery target. So he set the delivery standard for paying users to "delivery based on marriage outcome".

The "3-year" time limit was not arbitrarily decided. His team collected about 150 questionnaires from people waiting to get marriage certificates at the civil affairs bureaus in Nanshan and Futian districts of Shenzhen, to investigate how these newlyweds met, and how long they knew each other before getting married.

"Excluding natural relationships like classmates, colleagues, and childhood sweethearts, 70% of people who met through introductions or online platforms after starting their careers got married within three years," Zeng Xunxun explains: "If they were compatible, they got married; if not, they broke up. Three years is a very natural timeline."

Regarding concerns about users "delaying marriage for three years to maliciously get a refund", his judgment is: "For Chinese people, marriage is a major life event that involves far more expenses than 2000 yuan. Compared to the overall cost of getting married, 2000 yuan is not enough to influence people's decisions. Even choosing an auspicious wedding date costs more than that."

At present, MatchPerfect's product is still in its early stages. Whether the AI Matchmaker can consistently generate high-quality profiles, whether the accuracy of the dedicated matching model can reach 80%, and how many users will get married after three years — all of these outcomes need time to verify.

But in the practice of products like MatchPerfect, AI technology is not just solving social efficiency problems, it is also intervening in real social needs. There is a huge unmet demand for quality dating services among single adults in China. The previous generation of products only helped users "meet other people", but AI technology may