Lingguang 2.0 pivots to Personal Agent: Can Ant Group fill the gap in general AI entry points?

Text | Digital Intelligence Torrent, Author | Entropy Wild, Editor | Shen Xiao

When the Lingguang team appeared together on Xiaohongshu, it revealed Ant's determination to bet on Personal Agent.

After Lingguang 2.0 launched, the Lingguang team spoke continuously in the Xiaohongshu community.

The architect introduced Agent Harness, the product manager explained why Lingguang 2.0 should become "a personal assistant that can truly manage things," and the technical lead explained how to move AI from "being able to answer questions" to "being able to help you manage things."

The Lingguang team hopes that through community co-creation on Xiaohongshu, it can find the place in users' minds for this Personal Agent

product.

After all, Lingguang 1.0 was still a tool for generating mini-apps. By 2.0, Ant hopes users will hand things over to it. It needs to remember tasks, follow up with reminders, and become the user's personal secretary.

In the AI field, Ant Group's Afu and Abao have both performed well, but apart from these vertical products, Ant currently does not have a general-purpose AI product that, like Muse, Doubao, or Qwen, makes users think of it first when facing everyday needs.

Meta previously found itself in a similar situation. It had a social application ecosystem and a huge user base, yet it did not rank among the overseas giants in the AI contest.

Until Muse launched, reaching 10.2 million downloads in the U.S. market in just 40 days. It elevated Personal Agent to the AI pantheon and also gave Meta an AI card that the market could see.

Ant now needs its own Muse. Lingguang 2.0 is trying to take on this role. For Ant, "whether Lingguang can do it well" is not the core issue; getting into the game and trying in time is more important.

During the 2025 Spring Festival, DeepSeek went viral, and Ant held three consecutive days of review meetings starting from the first day back at work.

Ant Group CTO He Zhengyu later mentioned in a media interview that the team was excited, felt urgency, and even felt ashamed: Bailing had a similar starting point to DeepSeek in some model technologies, but was a step late due to reasons such as computing resources.

In March, Ant formed a relatively independent AGI project organization. By the time the Lingguang 1.0 product launched in November, this team had expanded to more than 200 people, directly led by He Zhengyu, with Cai Wei serving as Lingguang product lead.

At that time, the AI coding track had already produced a number of impressive growth examples.

For example, the Swedish startup Lovable announced on July 23, 2025, that its ARR had exceeded $100 million. It took only eight months to go from $1 million to over $100 million, users had created more than 10 million projects cumulatively on the platform, and about 100,000 new ones were added each day.

Keeping up with this wave of AI-generated applications, Lingguang officially launched its product on November 18, 2025, with the market slogan "generate a flash app in 30 seconds." Users do not need to understand code; they can customize a dedicated mini-tool by describing their needs orally.

He Zhengyu and Cai Wei admitted in a group interview at launch that when the direction was finalized in March, the team could not be sure whether flash apps could be delivered on schedule, and that doing this was itself a gamble.

Initial internal expectations were not high, with the target set at only 200,000 users by the end of the year. According to He Zhengyu, when Jack Ma saw this target, he thought it was too poor and asked the team to "push forward and charge for first place."

Lingguang indeed lived up to expectations. On the third day after launch, downloads exceeded 500,000, topping the App Store free tools chart. It broke 1 million in four days and exceeded 2 million on the sixth day. By comparison, ChatGPT had only 606,000 downloads in its first week, and Sora2 took five days to break 1 million.

But rapid short-term user growth did not translate into long-term success.

Four months later, in March 2026, data from the QbitAI think tank showed that Lingguang's three-day retention rate for new users was only 21.2%, compared with 30.92% for Doubao and 27.99% for Qwen over the same period, indicating that Lingguang's retention was clearly weak. QuestMobile data also showed that Lingguang had then fallen out of the top ten domestic AI applications by monthly active users.

On April 20, Lingguang continued to double down on flash apps, launching a 100 million yuan creator incentive plan: up to 10,000 quality apps rewarded per day, with 100 yuan per item; up to 100 creators rewarded per week, with a maximum of 10,000 yuan per person.

The prerequisite for receiving incentives was that works had to be publicly released in the newly launched "Lingguang Circle." Official data claimed that in the five months since launch, users had generated more than 30 million flash apps cumulatively.

Unfortunately, flash apps were like a second OpenClaw, and the hype quickly passed. In June 2026, market news said that Lingguang's "number one" Han Luo was transferred to support Afu, and several key technical staff were transferred along with him. Ant did not formally respond.

By autumn, another story attracted the market and gave Lingguang a second chance.

On September 8, Meta released Muse, directly stepping in to help users get things done: writing emails, planning itineraries, and negotiating bills.

According to SensorTower statistics, in the 40 days after Muse launched, downloads in the U.S. market alone reached 10.2 million. Its explosive popularity also pushed the concept of Personal Agent to the center of China's AI stage.

In less than two months, Lingguang 2.0 was officially released on October 27, with the product's main line fully switching from 1.0's flash app generation to the Personal Agent track. However, although it was billed as China's version of Muse, it failed to replicate the user experience expectations associated with Muse.

In actual use, users complain that there are blockers everywhere: users want it to recommend food delivery, but it cannot access Taobao Shangou's real-time store data; when asked to complete Taobao product price comparisons, it is limited by query interfaces that have not been opened; even for lightweight tasks within the Ant ecosystem such as collecting energy in Ant Forest or feeding chickens, it likewise lacks the corresponding operating permissions.

In 2024, Han Xinyi, who had just taken over as president of Ant Group, had his workdays packed with affairs, yet once the weekend arrived he instead fell into anxiety.

In an interview with LatePost, he admitted that he always had a matter weighing on his mind: general-purpose AI is reshaping the online traffic gateway, and once users hand their first sentence about booking tickets or shopping to AI, Alipay may very well end up handling only the final payment step, and never be opened throughout the entire process.

Ant must not only defend its payment base but also seize the first gateway where users initiate demands, and grasp the discourse power of AI to C.

The layout was launched very early.

In September of that year, at the Inclusion Bund Conference, Alipay officially released the standalone AI app Zhixiaobao, positioned as an AI life steward. The launch event demonstrated full-chain services such as ticket booking, food ordering, and ride-hailing, and the product also specially added a "Right Now" page that proactively pushes local services based on geographic location and user habits.

But the standalone app diversion approach did not become the final choice internally. The Alipay business team later embedded the AI steward capability directly into the main Alipay app, without creating a standalone diversion gateway—this is today's Abao.

Besides the life steward line, Ant also bet on the vertical AI assistant track during the same period. The health-oriented AI product AQ was launched in 2025, focusing on consultations, medication management, and health inquiries. Upon launch, thanks to the rigid-demand nature of medical scenarios, it quickly accumulated a batch of users.

Like Lingguang 1.0, AQ's development also drew Jack Ma's attention. In an interview, Han Xinyi once mentioned that Jack Ma hoped the product name would be easier to remember and more like a real person, reducing the psychological distance for ordinary people. On December 15, 2025, the health assistant AQ was officially renamed Afu.

Just one month after the rename, Ant spent hundreds of millions of yuan on advertising for Afu. For two consecutive years, 2024 and 2025, Ant's annual investment in the health AI track reached billions of yuan. At the same time, Lingguang's initial starting target of 200,000 users was also raised upward at Jack Ma's request.

In December 2025, Afu announced monthly active users exceeding 30 million and daily question volume exceeding 10 million. Ant finally had a to-C AI application with a considerable user scale.

On June 16, 2026, Abao opened invitation testing; it can be invoked by swiping right within the Alipay app, and the interface was streamlined into two major sections, Abao and Assets. When users book home appliance repairs, ride-hailing, food ordering, and other life needs, Abao can handle them within the Alipay system.

However, a person close to Ant told us that Lingguang's positioning within the group itself had not been fully finalized. Earlier this year, in an interview with LatePost, Han Xinyi also mentioned: Lingguang belongs to the "logic of experimentation," while Afu is the "logic of must-win," and the strategic weight of the two is not equal.

Subsequently, on September 21, Ant completed an important round of organizational adjustment: the Alipay Business Group, the Digital Payment Business Group, and the Zhima Credit Business Unit were merged into a new Alipay Business Group, headed by Wu Minzhi. Payment, merchant services, and credit capabilities were further consolidated to prepare for business implementation in the Agent era.

One month later, Lingguang 2.0 was officially released. For a new product still exploring its direction, mature businesses within the group tend to first observe its market feedback and then decide how much resources to open up, which better fits the interests of each business line.

Even within the internal ecosystem of a big tech company it is like this, and competition among major platforms is even fiercer.

After Muse went viral, domestic Personal Agent products quickly entered a concentrated launch period:

On September 24, the Chinese version of Today made its debut;

On September 28, Tencent's Lightweight Cloud team launched the cloud-based personal assistant LightVela;

On October 13, Qwen Go opened experience applications in the Qwen app;

On October 27, Lingguang 2.0 was officially released.

In addition, ByteDance is advancing a screen-automation Agent solution through the Doubao phone assistant, Baidu launched Dazi, and Tencent Yuanbao is ramping up personal errand capabilities.

When all the big tech companies see this opportunity, the Personal Agent itself becomes a contest of assembling coalitions.

Everyone wants to build a Chinese version of Muse and is unwilling to hand over permissions to an external AI steward, and the result is another round of platform wall wars: Amazon blocks Muse, 166 apps jointly refuse Doubao, just as Taobao once blocked Baidu's spider and WeChat blocked external links.

From this perspective, the victory or defeat of Personal Agent is not a contest of model capabilities, but a question of who can assemble a new AI friend-circle ecosystem that everyone is willing to join. In a situation where all parties guard against one another, the relatively neutral phone manufacturers in the middle of the chain seem naturally best suited to be this coalition assembler.

In June, Caijing verified that Honor was the first to obtain WeChat's A2A capability, and the phone assistant could initiate audio and video calls and send messages to designated friends; similar integrations by Huawei, Xiaomi, OPPO, and vivo were also in progress.

On August 17, vivo and Ant officially announced cooperation, with Blue Heart Little V linking up with Abao to connect services such as ride-hailing, invoicing, dining, and ticketing; on October 13, OPPO and Manus officially announced cooperation, integrating Manus Agent capabilities into the AI Mind Pilot of the Find X10 series.

Not only is the commercial position advantageous, but the technical characteristics of the Personal Agent implementation route further amplify the advantages of phone manufacturers.

When phone manufacturers and internet giants conduct targeted interface cooperation, the giants themselves decide which functions to open, and do not have to hand all operating permissions to the phone-side Agent. WeChat opening messaging capabilities to Honor and vivo linking with Alipay's Abao are both this model of cooperation.

On the commercial cost side, Morgan Stanley estimates that Muse costs about $10 to $18 per user per month to serve, of which the cloud sandbox virtual machine alone accounts for $5 to $8, far higher than ChatGPT's $2 to $3.

Even if Muse could achieve the same per-user monetization level as Facebook and Instagram, it still could not cover continuous computing power expenses.

But by cooperating with phone manufacturers, the Agent does not need to directly profit on its own, and the related computing power costs can be spread into the overall hardware cost. Phone manufacturers, in turn, gain an increase in hardware market share through the added selling point of AI errand capabilities.

In addition, the value of overseas Agents comes from the tedious manual communication in emails, independent websites, and scattered offline services. In China, however, shopping, payment, and local life have long been consolidated within super apps, and the operation chain has already been compressed to "two taps."

Users have no need, for the sake of a new technology narrative, to make simple operations complicated again, and they are even less likely to be willing to pay an extra token fee.

But a phone that works better because of cooperation with a big tech company's Personal Agent is a pretty good AI coalition terminal.