Source | Bohu Finance (bohuFN)

More than half of 2026 has passed, and Tencent's performance in the secondary market has not been optimistic.

Despite Tencent's best efforts to use its financial clout—accumulating HK$21.404 billion in share buybacks since 2026, making it the company with the most aggressive repurchases in the Hong Kong stock market this year—Tencent's stock price stood at HK$429 at the end of June, down nearly a third from its high of HK$633.7 at the beginning of the year.

The persistent decline in the secondary market is unrelated to Tencent's fundamentals. On the contrary, Tencent has been one of the internet giants with the strongest fundamentals in recent years, boasting a solid social media position, record-breaking gaming performance, and no distractions from food delivery wars or sudden new competitors. Nine years ago, when Tencent's stock price first touched HK$430 and its market capitalization surpassed HK$4 trillion, its adjusted net profit was RMB 65.1 billion. Now, although the stock price is similar, this figure has quadrupled to RMB 259.6 billion.

The main reason for the severe divergence between performance and stock price lies in AI.

As two of the dominant forces in Chinese internet, the differences between Alibaba and Tencent have led to vastly different performances in the AI wave. Alibaba has been more aggressive, rapidly transforming its company strategy and organizational structure. Its Qwen series has become the most influential model in the open-source community, with Hugging Face's top ten rankings once dominated by its derivative models.

However, Tencent seems to be stuck in its inertia (inertia) of catching up later and has not adapted to the pace of the AI era. After DeepSeek sparked a new AI narrative, Tencent Yuanbao quickly integrated the full-strength version and invested heavily in promotion, but the results were disappointing. Its self-developed Hunyuan large model has long lagged behind domestic peers, and its cloud services have been far surpassed by Volcano Engine and Alibaba Cloud.

At Tencent's shareholder meeting in May, Pony Ma described the situation of Tencent AI at the time: "A year ago, we thought we had boarded the ship, but later we found that the ship was leaking."

The hiring of Yao Shunyu, a former OpenAI researcher, was seen by the outside world as a turning point for Tencent AI.

After Yao Shunyu joined, Tencent's decision-makers provided significant support. By the end of last year, the Hunyuan team underwent a major reshuffle, with new leaders appointed for pre-training, post-training, evaluation, and Infra. Dedicated AI Infra and AI Data departments were also established to support Hunyuan.

In a previous conversation with Tang Daosheng, Tencent's Senior Executive Vice President, Yao Shunyu shared his understanding of AI: first, the foundational layer—how to make core technologies like pre-training and post-training robust enough; second, the product layer—how to truly implement technologies to create value for individuals and society; third, the frontier exploration layer—how to explore new research paradigms and industrial opportunities.

The Tencent Hunyuan Hy3, released on July 6, is clearly a product of this approach. Compared to its peers, Hy3 does not feature innovative architectural designs or focus on parameter scale. It uses the most standard MoE Transformer architecture, with a total of only 295B parameters. Each inference activates the Top-8, totaling 21B parameters, and supports a context length of 256K.

In addition to the model, Tencent has also significantly accelerated its AI product development. On March 9, WorkBuddy was officially launched; on June 5, at the Tencent Cloud AI Industry Application Conference, Tencent officially released the WorkBuddy enterprise version and the office intelligent agent suite, Agent Suite; in June, WeChat conducted a small-scale internal test of the AI assistant "Xiaowei."

On July 8, Tencent also announced that another former OpenAI researcher, Tian Yonglong, had joined the Large Language Model Department to participate in the research and development of visual language models (VLMs). Tian Yonglong and Yao Shunyu are undergraduate alumni from Tsinghua University and were core researchers who worked together at OpenAI.

So, the question arises: Has Tencent AI found the right direction this time?

01 A New Model: Tencent Takes a Seat at Its Own Table

Generally speaking, the parameter scale determines the upper limit of a model's intelligence. Nowadays, the parameter scales of most new models are in the trillions. For example, Xiaomi's MiMo-V2 Pro has a total of 1.2 trillion parameters, DeepSeek V4 has 1.6 trillion, and the recently released Kimi K3 has a staggering 2.5 trillion parameters.

The limitation of parameter scale destined (dooms) Hy3 to underperform in very complex tasks.

For instance, in the current hottest code track (arena), Hy3's performance is clearly unsatisfactory. On the SWE-bench Pro benchmark, Hy3 scored 57.9, significantly lower than Claude Opus 4.8's 69.2 and GLM 5.2's 62.1. On the SWE-bench Multilingual benchmark, it scored 75.8, also trailing Claude and GLM 5.2. On the Terminal Bench 2.1 terminal command benchmark, it scored 71.7, showing a clear gap from the top three.

LatePost also reported a small detail: the previously released Hy3 Preview underwent a round of internal testing within Tencent. During the process, multiple external models ran smoothly, but Hunyuan encountered issues.

Instead of taking a risky approach, Hy3 chose the safest technical route, partly because Tencent AI has too much historical debt.

According to LatePost, when the Hunyuan project was initiated in 2023, the project team did not even have a single GPU and had to borrow 2,000 from the advertising department. The core processes of large models are not secret; the key is to lay a solid foundation in infrastructure and data. Issues such as Infra's inherent deficiencies, chaotic data annotation, and missing training links prevented Hunyuan from establishing a solid foundation for building a good model. In comparison, Anthropic's co-founder, Jared Kaplan, is still reported to personally oversee data with his team every day.

Therefore, Hy3 focused its main efforts on these foundational aspects: redefining data standards and cleaning existing data from scratch.

Of course, another important reason is Yao Shunyu's judgment: in China, building a relatively small model that can match the performance of large models in most tasks and has strong robustness may be more valuable than improving by one or two points in some long, complex Fancy tasks.

From the results, Hy3 has largely met this expectation. Tencent's internal evaluation data shows that in a blind test involving real work tasks with 270 internal experts, Hy3 averaged a score of 2.67, higher than GLM-5.1's 2.51. The leading projects mainly focused on front-end development, data and storage, CI/CD, and other work tasks. Among them, the hallucination rate of the official Hy3 version dropped from 12.5% to 5.4%, the common-sense error rate dropped from 25.4% to 12.7%, and the multi-round dialogue problem rate dropped from 17.4% to 7.9%.

From a reasoning cost perspective, each inference activates only 21B parameters, with a reasoning cost about one-seventh of GLM-5.2's. In terms of API pricing, Hy3 is priced at RMB 1 per million tokens for input and RMB 4 per million tokens for output, with a cached input price of RMB 0.25 per million tokens. This pricing is significantly lower than other flagship models.

With a total of 295B parameters and only 21B activated, Hy3 became Tencent's first model to top OpenRouter's weekly usage rankings, exceeding internal expectations.

Hy3 is not a stunning model, nor is it one that can propel Tencent into the top tier of domestic AI. However, it practically conveys that Tencent has returned to the right track of AI research and development.

02 The Turning Point of the Second Half

Entering 2026, although the vision of AGI remains, how to make AI truly create value has become a question that all practitioners need to consider.

On the one hand, the shift in AI narratives is already evident.

Earlier this year, openclaw ignited token consumption, and internet giants turned token usage into a leaderboard, with everyone installing "lobsters" on their computers.

But the trend changed quickly. Soaring costs prompted major companies to slam on the brakes for AI usage. Uber directly set a monthly token usage limit of $1,500 per person. Meta issued a memo to 6,000 core employees, clarifying token quotas for all staff. Amazon executives also publicly warned employees not to "use AI for the sake of using AI," shifting performance metrics from token consumption to standardized business delivery outcomes.

"Token inefficiency" has forced companies to return to business fundamentals and begin cautiously evaluating the actual return on every dollar invested in AI. Usability has become the most important standard.

On the other hand, the market has shown a clear K-shaped divergence: the prices of general-purpose large models continue to decline, trending toward "infrastructure," while high-end models with complex reasoning capabilities may maintain a premium.

This year, the domestic large model API market has experienced an unprecedented wave of price cuts driven by "technological cost reductions." For example, DeepSeek's flagship model, V4-Pro, announced a permanent 75% price reduction, bringing the overall price down to 25% of the original. Following DeepSeek, Xiaomi announced a permanent price reduction for its MiMo-V2.5 series APIs, with some scenarios seeing reductions of up to 99%. Other domestic manufacturers have also followed suit with price cuts.

Price is becoming a crucial aspect of large models. The Information reported that DeepSeek charges only a fraction of the price of OpenAI and Anthropic's similar models for access to its latest flagship model, V4, while still maintaining a gross profit margin of over 50%.

From this perspective, if the entry point of the PC era was search and the entry point of the mobile internet era was social media and short videos, then in the AI era, the most important aspect of applications is what value they can provide to users. Compared to its peers, Tencent has the most comprehensive set of assets, including applications, traffic, and scenarios. As long as the model's capabilities and costs are up to par, token consumption will quickly increase.

The birth of Hy3 has filled an important gap in Tencent's productivity scenarios.

On July 6, WorkBuddy and Yuanbao App were fully integrated with Hy3. The next day, Tencent's self-selected stocks were fully integrated. Now, dozens of products, including ima, Marvis, QQ Browser, WeChat Reading, and WeGame, have all integrated Hy3.

With Hy3 and WorkBuddy at its core, Tencent has effectively formed a product pathway for general productivity scenarios. Tencent revealed that since the launch of Hy3 preview, its daily average token consumption has increased 20-fold.

Based on Hy3, Yuanbao also simultaneously launched the Agent function. In its internal evaluation, Hy3 has surpassed domestic models like GLM 5.1 in comprehensive office and life service scenarios, sufficiently supporting real business workflows. Users can input their needs in daily conversations, and Yuanbao can directly execute complex tasks and deliver files such as PPT, Word, Excel, PDF, and HTML to meet daily office demands, all for free.

In addition, Hy3 is also helping other internal Tencent businesses transform. For example, Peacekeeper Elite recently launched a new AI teammate, Xiaotian, supported by Hy3, upgrading multiple functions. WeGame's recently launched Path of Exile: Ascendancy AI game assistant, after integrating Hy3, saw its comprehensive success rate for multi-round reasoning and tool scheduling increase to 92%, with the hallucination rate dropping from 4.5% to 2.8%, significantly improving output accuracy.

Hy3 is becoming the key to connecting the ecosystem, and a large amount of feedback from scenarios and users provides valuable data for model iteration.

03 In Conclusion

Apart from models and products, Tencent has not fallen behind in its AI investments. Domestic large model startups such as Minimax, Baichuan Intelligence, Yuezhiaimian, and Zhipu all have Tencent's involvement. Tencent has also invested in industrial chain (supply chain) companies like Suiyuan Technology and Moore Threads.

This year, Tencent has invested in DeepSeek and Kling. There are also reports that Tencent is forming a Chinese capital consortium to negotiate the repurchase of all shares of Manus from Meta at an estimated valuation of $2 billion. Manus is the world's first truly viable general-purpose Agent product.

Large companies have more options than startups, but this also means more departmental silos, power struggles, and a lack of an all-in commitment. AI has changed traditional organizational forms and talent relationships, and large companies have not yet fully adapted.

Hy3 has made a good start, but the real challenges clearly lie ahead.

Referenced Sources:

1. Overseas Unicorn: Deconstructing Anthropic: The Best AI Company May Also Be an Organizational Invention

2. Geek Park: Tang Daosheng in Conversation with Yao Shunyu: Is Tencent AI Slow?

3. Tencent Research Institute: Token Inefficiency

4. LatePost: When a Young Talent Drops In: 300 Days Transforming Tencent Hunyuan

5. Narrowcast AI: Hy3 with WorkBuddy: Tencent Accelerates Entry into Productivity AI | Narrowcast Weekly

6. InfoQ: Tencent Hunyuan Hy3 Officially Released, Yuanbao Simultaneously Launches Hy3 Agent Capabilities for Free

7. Tech Planet: Tencent Acquires Manus for $13.5 Billion, Going on an AI Shopping Spree

8. Yiketalks: Behind Tencent's $2 Trillion Evaporation: Pony Ma's Ship, Yao Shunyu's Oars, and the Cruelest Valuation Logic in the AI Era | Hardcore Observations

The cover image and accompanying illustrations are owned by their respective copyright holders. If the copyright holders believe that their works are not suitable for public browsing or should not be used free of charge, please contact us promptly, and our platform will make corrections immediately.