Patience is a virtue, and haste often leads to waste.

Article | Huashang Taolue by Fu Shuai

Beijing, February 17, 2025.

A high-profile symposium for private enterprises convened in the capital, drawing several renowned entrepreneurs, including Ren Zhengfei and Jack Ma.

Video footage from the event clearly shows Liang Wenfeng, founder of DeepSeek, and Pony Ma, founder of Tencent, sitting side by side, with their companies both embarking on new AI-centric narratives.

▲Image source: CCTV News screenshot

【From Dormancy to Catching Up】

On January 26, DeepSeek-R1 gained domestic popularity after initially attracting attention overseas.

After using it, Feng Ji, producer of "Black Myth: Wukong," couldn't resist praising it: "DeepSeek is a national-level technological achievement. It's hoped that AI will become as ubiquitous in our lives as water and electricity."

China's fiercely competitive landscape for large AI models has become even more intricate and complex.

2024 marked the inaugural year of large AI models in China. Not only did numerous startups, including the AI Six Tigers, make continuous moves, but major internet giants also personally entered the fray, joining the melee.

Among them, ByteDance leveraged its platforms such as Douyin and Toutiao to aggressively invest in traffic, pushing the download volume of Doubao APP to the top of the domestic AI application rankings, with monthly active users exceeding 50 million.

Meanwhile, Baidu's founder, Li Yanhong, frequently appeared to promote his company's Wenxin Yiyan model.

Alibaba's self-developed Qwen series of large models became the most popular open-source large model in China and is also leading internationally.

In contrast, Tencent, as the internet giant with the highest market capitalization in China, seems to be lagging behind in the AI era. While competitors frequently showcase their strengths, Tencent has been slow to produce impressive AI products, even being criticized by the outside world for "always being half a step behind in AI layout."

In fact, Tencent has not invested little in AI over the years. The reason why the outside world has not perceived much movement is that Tencent itself does not like to make a fuss, and its products indeed lack a splash in the C-end market.

According to Tencent's 2024 financial report released on March 19, the company's R&D investment in AI reached 70.69 billion yuan over the past year, with cumulative investment over the past seven years amounting to 391.2 billion yuan, a significant investment.

As early as 2023, Tencent launched its self-developed Hunyuan large model, adopting the Mixture of Experts (MoE) architecture later popularized by DeepSeek, which can significantly save computing resources.

Moreover, the flagship model of the Hunyuan series of large models had reached a trillion-parameter scale at that time, with its capabilities firmly ranked in the first tier domestically. However, it is a pity that Tencent seemed uncertain about the direction of AI, so it did not vigorously promote it, and the perception from the outside world was not strong.

In a way, the popularity of DeepSeek-R1 accelerated Tencent's AI strategy, and it was like a tiger waiting for its prey for a long time, suddenly moving quickly.

First, it resolutely adopted a multi-model strategy of "self-development + open source," successively integrating DeepSeek-R1 into products such as Tencent Yuanbao, forming a dual-engine architecture of "Hunyuan + DeepSeek" to accelerate AI penetration across all products.

With the momentum of DeepSeek-R1, more people began to learn about Tencent's AI products, and Tencent also started intensifying the development of its own products.

On March 21, Tencent launched the official version of its self-developed deep thinking model Hunyuan T1. Built on Turbo S, the industry's first ultra-large-scale Hybrid-Transformer-Mamba MoE large model released by Tencent in early March, it is the first time in the industry that the hybrid Mamba architecture has been applied losslessly to ultra-large-scale inference models.

This architecture can effectively reduce the computational complexity of the traditional Transformer structure of large models, significantly improving inference speed. In actual operation, its articulation speed can reach up to 80 tokens/s, faster than DeepSeek-R1.

In addition, according to evaluation results from multiple public datasets such as MMLU-pro, CEval, AIME, and Zebra Logic, T1 has reached the level of industry-leading inference models in terms of Chinese and English knowledge, competition-level mathematics, and logical reasoning metrics.

At the same time, Tencent also synchronized adjustments to the group's organizational structure to match its strategic shift.

Starting in February this year, the teams of core To C products, including QQ Browser, Sogou Input Method, AI Workbench ima, and AI Assistant Yuanbao, have been transferred from PCG (Platform and Content Group) to CSIG (Cloud and Smart Industries Group), further accelerating the deep integration of the technology middle platform with industrial scenarios to provide users with higher-quality AI experiences.

Tencent Yuanbao has also overtaken ByteDance's Doubao, which was in the spotlight last year, to become the new leader in domestic AI applications.

【National-Level APP Meets National-Fortune-Level AI】

Although many Tencent applications have integrated DeepSeek-R1, in terms of the shock and influence among ordinary users, WeChat is undoubtedly the most significant.

For the general public, they may not necessarily use ByteDance's Doubao, Tencent's Yuanbao, or Baidu's Wenxin Yiyan, and they may not even know what an AI assistant is. However, there is a high probability that they have WeChat on their phones.

As a national-level APP with over 1.3 billion monthly active users, WeChat's user base is unmatched in China.

Thanks to WeChat, Tencent's vast array of businesses can connect and activate each other. For this reason, every move of WeChat will be infinitely amplified, attracting millions of attentions.

WeChat is Tencent's biggest trump card, and the combination of "WeChat + DeepSeek" may unleash energy that other third-party applications cannot achieve, becoming the key to Tencent's realization of the To C AI strategy.

Regarding the current implementation of global large AI models, search is the largest application scenario. However, today's search landscape is no longer dominated by search engines like Google and Baidu.

The rise of mobile internet has encapsulated numerous data points into independent APP applications, causing the internet world to split into numerous "data islands."

Zhihu can do search, Xiaohongshu can do search, and so can WeChat.

As the current gathering place for the highest-quality content on the Chinese internet, WeChat official accounts possess a vast amount of high-quality Chinese information. "WeChat + DeepSeek" will create a winning combination of "the best content search + the strongest AI model."

Other platforms are difficult to reach such heights.

The excitement doesn't stop there.

What is WeChat? Besides being a chatting tool and a content gathering place, it is also a natural application community. Countless third-party mini-programs are gathered within it, providing us with various convenient services - scanning codes to order food, booking tickets in advance, checking flight information, etc.

In the future, with AI as a lever, this advantage may be further amplified.

For example, many people have tried using large AI models to create a customized travel plan before traveling. The resulting content may seem clear and well-reasoned at first glance, but upon closer inspection, it does not hold up to scrutiny.

This is because when AI cannot obtain complete and accurate information, it will only "make up stories" according to predetermined logic to try to "perfectly" complete the task you assigned to it, leading to serious large model hallucinations.

However, WeChat integrates a wealth of third-party mini-programs, including official mini-programs of tourist attractions across the country.

Once their API interfaces can be connected with large AI models, similar problems may become much less frequent. This "search-as-a-service" model will greatly promote the implementation and penetration of AI.

By then, WeChat will not just be a social platform but will evolve into an intelligent agent platform, helping you automatically perform tasks such as booking flights, writing documents, and analyzing financial reports.

The idea of freeing up your hands is delightful.

When a national-level APP meets national-fortune-level AI, a brand-new and truly practical AI Agent may already be on its way.

【Plan Before You Act】

Why was Tencent, one of the most powerful players, not so high-profile or aggressive towards AI before 2025?

Was it lax? Did it fall behind?

Now, it seems that it did not.

On the one hand, it was quietly accelerating its AI layout, besides AI-izing its own business and developing AI projects, such as introducing AI into more advertising, marketing, and enterprise service scenarios, training basic models, and also matching investments.

For example, it has successively invested in leading large model enterprises such as Zhipu AI, Baichuan Intelligence, and Dark Side of the Moon.

On the other hand, it was also observing, thinking, deliberately slowing down its pace, and waiting for an opportunity to "hit the bullseye" with one strike.

At Tencent's 2023 shareholders' meeting, when ChatGPT was booming globally and AI became a concern for many investors, while domestic giants were already rolling up their sleeves and working hard, Tencent seemed unmoved. Someone asked Pony Ma a question:

What is Tencent's view on AI and how does it respond?

Pony Ma's answer was once considered an excuse for slow action.

He said, "Many internet companies have accumulated a lot of experience in AI and are doing related work. Tencent is also working hard on research and development, but it is not in a rush to complete it early and showcase half-baked products.

'AI is an opportunity like an industrial revolution that comes along once every few hundred years, similar to the invention of electricity. In such a major opportunity, bringing out the light bulb (AI) one month earlier is not that important.'"

At Tencent's earnings call on March 19 this year, after Tencent Yuanbao had come from behind, Pony Ma once again talked about AI-related content. His words seemed to echo his previous impatience:

"After careful consideration, Tencent reorganized its AI team a few months ago to focus on rapid product innovation and in-depth model development. Although AI expenditures currently account for a low percentage of its revenue, based on the opportunity for applications development, Tencent's revenue growth rate last year will continue to increase, and this amount of capital expenditures in AI will nearly double to 100 billion yuan."

Although this investment is slightly lower than the annual investment amount previously announced by another internet giant, Alibaba, the difference is not too significant.

Although the real war has just begun, looking back at the past two years, from being considered lagging behind to suddenly taking the lead now, it seems to confirm Pony Ma's initial judgment on the development of AI - no need to rush, patience is a virtue, plan before you act, and it will be in time.

This is a strategy and a kind of composure, which is worth learning from.

【References】

[1] "Tencent to Invest Hundreds of Billions in AI", Quan Tian Hou Ke Ji

[2] "Tencent is Still Tencent", Xin Xi Ping Quan

[3] "Pony Ma: AI is an Opportunity Like an Industrial Revolution That Comes Once Every Few Hundred Years, Similar to the Invention of Electricity", Zhong Guo Ji Jin Bao

——END——

Welcome to follow [Huashang Taolue] to recognize influential figures and read tales of strategy.

All rights reserved. No unauthorized reprinting is allowed.

Some images are sourced from the internet.

If there is any infringement, please contact us for deletion.