After a few months of dreamy starts, Tencent Yuanbao finally returned to its cruel real world.
The battle for the large model list that lasted for several months has come to an end. As of May 20, the first place in the Apple App Store free App download list was Doubao, DeepSeek ranked sixth, and Yuanbao had fallen to 43rd. In the latest quarterly financial report, Tencent did not pay too much attention to Yuanbao, and its core is still upgrading the game and WeChat ecosystem. In the conference call after the earnings report, when asked what more collaboration Yuanbao will have with WeChat, Ma Huateng’s answer was, “Frankly, it is still in a very early stage.”
Management’s avoidance of this topic is an attitude. This shows that the “ingots” that have been pushed to the front of the stage may not be able to carry the banner of Tencent AI.
AI entrepreneur Ah Can told NoNoise Reduction that the core of AI tools is the cooperation of models and applications, but the combination of Tencent Hunyuan and Yuanbao does not seem to be ideal. “When Yuanbao accessed WeChat, I also thought it might work, but after using it a few times, the ability to follow the instructions was so poor that I wondered if Tencent did it”, “Many of them are copied from others, such as the product of KPIs.”
One user also expressed this confusion to us. Although the mobile phone has always been equipped with ingots, it does not have a strong sense of presence, “it seems that it is okay to have it or not”.
At the 2023 Tencent shareholders’ meeting, Ma Huateng emphasized that Internet companies have a lot of accumulation in the field of AI, and Tencent is also immersed in research and development, but it is not in a hurry to show semi-finished products early, “The key is to do a solid job in the underlying algorithms, computing power and data, and more importantly, the scenario is implemented, and we are still doing some thinking.”
He believes that some companies are now too eager to boost stock prices, “We have never been in this style.”
In just two years, the C position of the large model track has changed one stubble after another, but the protagonist has never been Tencent. It is difficult to say whether Tencent can still calm down at the moment, as can be seen from the recent public statements of Ma Huateng and Tencent’s management, Tencent’s AI acceleration is inevitable.
Yuanbao is difficult to bear heavy responsibilities
The model ability is poor, and you can feel the gap after a few uses.” AI quantitative practitioner Cong Fei expressed his disappointment with Yuanbao to us. In fact, Yuanbao is not “working hard”. After transferring from TEG (Technology Engineering Business Group) to CSIG (Tencent Cloud and Industry Business Group) in January this year, and separating from the hybrid large model technology research and development system, Tencent Yuanbao’s iteration has accelerated, especially after accessing the full-blooded version of DeepSeek-R1, which has been updated more than 30 times in March this year alone. Helplessly, it takes time for AI products to go from “rough” to “hardcover”, just like a poor student being sent to a national-level famous teacher, his grades cannot improve by leaps and bounds within a few days, and make-up classes still have to start from smashing the foundation. Previously, the Buddhist Tencent Yuanbao also had similar problems. Before accessing DeepSeek, its product perfection lagged behind its peers. In mid-March, a Tencent shareholder also summarized more than 10 product suggestions from the community of various platforms into more than 10 product suggestions and fed them back to the Tencent Yuanbao team. These suggestions include: total errors in compound interest calculations; There are many problems with large model hallucinations that need to be cross-confirmed by multiple models; Yuanbao PC is too rudimentary, and I hope to integrate with ima to launch a PC browser as soon as possible. Ma Huateng replied personally: “Many of them are in the planning and have very new ideas. 」
On April 16, Yuanbao launched a new feature that allows you to add it as a friend and talk directly in WeChat. Like other products such as ChatGPT, Yuanbao is embedded in the WeChat ecosystem as an entry point to lower the threshold for use.
It is said that Yuanbao, which is connected to the Hunyuan T1 model and DeepSeek-R1 at the same time, should be more convenient in reasoning and calling the WeChat system, but for users, the feeling does not seem obvious. “Yuanbao’s DeepSeek R1 was the first time I encountered a situation where the model ended deep thinking without thorough reasoning, and then there was no conclusion. Although it is a bit of a complex reasoning, it is by no means a super difficult problem, at least it can be completely derived by switching to other platforms such as Volcano Engine’s DeepSeek.” The above user said that this made him suspect that Yuanbao’s DeepSeek did not know whether it was a secretly castrated version or limited parameters.
▲Netizens on social platforms questioned the Yuanbao version of DeepSeek
Throughout April, giants including Alibaba Cloud, Baidu, and Byte have made great progress and upgraded their respective large models. In addition, DeepSeek-R2 is rumored to be launched in May, which has cast a layer of uncertainty on this year’s large model battle. Tencent, which has internal and external troubles, finally has a sense of urgency. On April 29, Tencent announced the establishment of a new large language model department and a multimodal model department for Hunyuan, and at the same time reorganized data platforms, machine learning and other departments to provide computing power and data support for Hunyuan. According to media reports, Tencent Data Learning and AI La have also conducted “horse racing” for the dominance of the mixed meta, and some independent small models have been built on each line, which eventually led to the waste of resources. This upgrade of Hunyuan also means the end of the inefficient horse racing mechanism in the past. Tencent has clarified the strategic priorities driven by hybrid technology, rather than relying solely on business-driven.
However, Cong Fei believes that the current situation of Yuanbao being both an entrance and a depth in the WeChat ecosystem is destined to not coexist for a long time. From a common sense point of view, Tencent should hope that more people will call the hybrid model after Yuanbao is connected to WeChat, train and accumulate data, and make deep binding with the WeChat ecosystem. But the real situation now is that everyone only wants to use DeepSeek through Yuanbao, because it is more convenient, and Yuanbao also calls DeepSeek by default, resulting in relatively few users who actually use Hunyuan. “I remember that before the advent of the smartphone era, Nokia also engaged in a dual-system strategy – both retaining Saipan and developing Meego. The result was like chasing two rabbits at the same time, and in the end I didn’t even catch one.”
Sometimes, although the low-hanging fruit is easy to pick, the taste may not be good.
The dilemma of throwing into the water
Although as early as 2022, Ma Huateng mentioned at the internal conference that “I no longer believe in the story of buying volume, and we will not do anything about buying volume in the future”, but after fully connecting to DeepSeek on February 13 this year, Tencent once again spent money to invest in the flow and made every effort to support ingots. On March 3, Tencent Yuanbao App surpassed DeepSeek for the first time and topped the free app list of China’s Apple Store. Just 6 days later, Yuanbao’s ranking slipped to third place behind DeepSeek and Doubao.
According to data, in February 2025 alone, Yuanbao’s investment amount will be as high as 303 million yuan, nearly ten times the investment before accessing DeepSeek, and the investment in March 2025 will be close to 1 billion yuan. Huge investment is at the heart of this round of AI arms race. At the beginning of this year, Alibaba also announced that it would invest more than 380 billion yuan in cloud and AI hardware infrastructure construction in the next three years. In the fourth quarter of 2024, Tencent’s capital expenditure on AI increased to 39 billion yuan, which not only exceeded the total expenditure in the first three quarters of 2024, but also exceeded the annual expenditure in 2023. Previously, some industry insiders said that the retention rate of AI products is much lower than that of mobile Internet products, and the weekly retention rate of DeepSeek, Doubao, and Yuanbao is only about 25%, which inferres that its 30-day retention may be less than 2%. In the face of highly homogeneous competition, there seems to be no better solution, and the only thing they can do is to maintain superficial “prosperity” by throwing money into flow. Behind the excitement is the problem of retention that cannot be avoided.
▲Yuanbao’s advertising in the circle of friends
In the conference call after the earnings report, Tencent President Martin Lau admitted that Yuanbao does have a certain amount of investment, but in the future, more attention will be paid to the retention rate and activity rate, and more importantly, the organic linkage and collaboration between Yuanbao and other products. Behind the retention problem is actually the problem of product strength, no matter how complex the deduction, for users, ease of use is the highest affirmation. In the view of the above-mentioned people, regardless of Doubao, Yuanbao, DeepSeek, there is still a gap between the underlying strength and ChatGPT and Gemimi, etc., and it is necessary to accelerate the evolution. If it weren’t for the emergence of DeepSeek, Tencent might still focus on “defense” in this battle of AI. As for when to show the sword, I am afraid we will have to wait for the really suitable scene to appear. “AI currently does not have a positive cycle business model, and it seems that there is no threshold. The real moat is resources, channels, and entrances. Tencent has no shortage of these things, and it is not too late to realize it after the market has an effective path. Tencent still wants to pick peaches directly,” Ah Can said.
Tencent is “overrated”
Tencent wants to obtain AI fruits through WeChat, but the increasingly heavy WeChat may not be able to carry all of Tencent’s ambitions. People’s concerns about Yuanbao and Tencent AI mainly come from the fact that many of Tencent’s products have staged a script of “high opening and low going” in recent years. Last year, “Noise Reduction NoNoise” reported on the “gift-giving” function of WeChat stores. At the beginning of the launch, the outside world believed that this would be an exploration of the same magnitude as WeChat red envelopes. However, judging from the recent volume, the gift-giving of small stores has not replicated the miracle of WeChat red envelopes. This has also triggered a re-examination of WeChat social e-commerce. Many people are gradually realizing that WeChat is not omnipotent. WeChat store was launched as early as 2020, WeChat hopes to open up the ecology of video accounts, mini programs, search and search, but the scale of e-commerce orders is not large, and the monetization mainly depends on advertising business income. In August 2024, WeChat upgraded the Channels store to a WeChat store, hoping to seek more conversions in WeChat scenarios.
▲Gift button on the product page of the WeChat store
The latest development is that WeChat has established an e-commerce product department. Just when people’s enthusiasm for WeChat e-commerce was rekindled, Martin Lau took the initiative to pour cold water. He said that this is only because of the split and independence carried out by the expansion of the scale, not a major adjustment, and there is no need to over-interpret it.
If the pace of innovation within WeChat is not satisfactory, then Tencent’s performance in some technical products outside of WeChat in recent years has also been lackluster.
For example, Tencent Docs, which was born in 2018, is probably the most evaluated by users as “difficult to use”. Against the backdrop of Feishu’s amazing feeling, Tencent Docs’ functionality and experience are difficult to match.
In Zhihu’s last discussion of Tencent Documents, some users attributed the defeat of documents and other products to Tencent’s “arrogance”. “When I contacted the product manager with the crash, they replied, ‘Our other users haven’t experienced this.'”
Learning the lessons of “arrogance” in the past, in this round of AI competition, Tencent decided to put down its position. Previously, according to the “Latepost” report, when Alibaba and Byte were fully attacking the large model hardware, Ma Huateng once told the AI team, “We must cooperate well with the outside world and don’t think about doing everything ourselves.” He also said, “Be soberly aware of the actual situation and don’t overestimate your abilities.”
Under humility, Tencent really made an amazing AI-native product – intelligent knowledge base product “ima”. The product has received rave reviews in the circle, but it has not yet broken the circle on a large scale. After all, only Yuanbao has the investment treatment of the group’s strength.
▲Knowledge base creators can publicly publish knowledge bases on IMA
According to the latest quarterly financial report, Tencent’s Q1 revenue reached 180 billion yuan. Although the growth rate hit a new high in more than three years, profit growth began to slow down. Ma Huateng said in the conference call: “In the first quarter of 2025, AI capabilities have made substantial contributions to businesses such as performance advertising and evergreen games, and have also increased investment in opportunities such as Yuanbao applications and AI in WeChat.” This also confirms that the market has a consensus on the advertising value brought by AI, and as for others, it seems to be still at the level of imagination. “Tencent’s current state must be left behind, not because it is not without effort, but after waiting for more scenarios and commercialization of large models to mature.” Cong Fei believes that the current swing is more like Tencent’s accumulation. After all, other AI products have fewer application scenarios and weaker moats.
At the earnings call, Tencent’s management clearly released the company’s long-term AI strategic intention: to create a “unique agent” that is deeply bound to the WeChat ecosystem – it can connect the WeChat ecosystem and millions of mini programs. These mini programs cover a wide range of information and have the ability to trade and operate across multiple applications. Compared with general-purpose Agent AI, this is a differentiated product that only Tencent can make. As a national-level App, WeChat is still Tencent’s biggest certainty. But you can’t rely on a card on WeChat for everything, right? (At the request of the interviewee, the text is pseudonym)
