Source | Bohu Finance (bohuFN)

In just one week, Tencent pushed its hidden AI cards much further forward.

On October 29, Tencent first brought out Buddy One at the application layer, using WorkBuddy as the main workspace and linking Tencent Meeting, Tencent Docs, Tencent Drive, and ima into one line; then on November 4, the underlying Hunyuan large model delivered the official Hy4 version, concentrating on filling the shortcomings that had long been talked about, including visual understanding, computer operation, and long-horizon tasks.

From the underlying foundation to the upper-layer applications, the gears of Tencent's models and products have finally meshed together, and it can be said to have taken its seat at the table properly with a full set of chips. For Tencent, which has been repeatedly questioned over the past two years for being a beat slow, this is a rare direct response.

By all accounts, Tencent should have reached a point where it could breathe a sigh of relief. The problem is that scrutiny from the market has also changed during this period.

Just a week before Tencent played its cards intensively, Google's parent company Alphabet released its financial report. Cloud revenue growing 103% year over year in the quarter did not bring the expected investor enthusiasm, and on the first trading day after the report, Alphabet's stock price fell 12.64%.

Investors no longer listen only to technology and product stories, but have begun calculating the return on investment of companies' AI spending. Evolved to this day, it has solemnly become a stock-market-crash-like violent selloff targeting large tech stocks, and any AI-first tech company needs to reassess the input and return of AI infrastructure.

At this time, Tencent can barely sit down and seriously talk about its AI prospects. The question this giant needs to answer is much thornier than simply training a good model or assembling a good product.

01 The foundation model takes the table

The progress of the official Hy4 version is first reflected in that it has truly expanded its capability boundary to the level of doing work.

This model, using a Mixture of Experts (MoE) architecture, has total parameters pulled up to 770B, activated parameters of 49B, and a context window breaking through 1 million Tokens.

Compared with the Hy4 preview that explored the path at the end of August, the official version not only continued to increase investment in pre-training and post-training, but also for the first time packed in visual understanding and computer operation capabilities. It can read and understand complex professional charts, and can also combine command lines with graphical interfaces to directly operate professional software such as Blender and CAD, completing continuous cross-application tasks in one go.

In the financial public test initiated by Tencent together with a fintech team, 213 industry experts scored 579 real business tasks, and Hy4's overall performance surpassed the open-source models of DeepSeek, Kimi, and GLM from the same period; in its own WorkBuddy actual test, its task pass rate reached 96.15%.

Although these results carry the black box of vendor self-testing, and the official version's model weights have yet to be fully released, Hunyuan has indeed accumulated enough qualification to participate in competition at the open-source forefront.

After all, Tencent's chief AI scientist Yao Shunyu has explicitly complained that there is a not-so-good tendency in the domestic industry, namely being overly keen on chasing leaderboard rankings. He cares more about how to set the right Eval (evaluation standards) internally and pursuing Co-Design between models and products.

It can be understood this way: any Tencent model progress is left to users to inspect performance in actual tasks, rather than looking at Benchmarks. What the outside world should pay more attention to is whether model iteration has achieved a sufficiently agile rhythm.

Tencent's answer did not disappoint the market. Since Hy3 preview explored the path in April, Tencent has been running in small quick steps all the way, launching the official Hy3 version in July, handing over to Hy4 preview at the end of August, and by November, the official Hy4 version landed as scheduled.

The small quick steps move fast, but the approach has always maintained a similar style. Whenever a new model is open-sourced, Tencent always prioritizes connecting the model to its own products such as Workbuddy, first putting the preview version into its own products to collect real problems, and then using the official version to focus on fixing capability shortcomings.

Developing this set of small combo moves is already not easy for a major company previously often criticized by the outside world for being bloated in organization and half a beat slow in action, and it carries more weight than a one-time concentrated computing power effort to stage a leaderboard sprint.

Tencent thereby proves that it is not only able to participate indirectly in the AI race by buying, relying on massive traffic, and existing scenarios; it is also capable of delivering the achievements of underlying models into real workflows frequently and stably.

It should be noted that Tencent Hunyuan also stated frankly in its technical blog that Hy4 still has a relatively large gap with the world's top closed-source models, especially on hard nuts such as frontier mathematics, long-horizon development, model self-iteration, and complex visual question answering.

From this perspective, Hy4 is merely a phased answer sheet.

And to understand why Tencent has been able to adjust its breathing and keep trotting continuously over the past six months, one must shift their gaze behind the model and see how it actually performed that organizational surgery on itself.

02 The reserve force of the AGI organization

Tencent's overall changes need to start from a larger goal of Yao Shunyu.

In June this year, in a conversation on the same stage with Tencent Cloud CEO Tang Daosheng, Yao Shunyu gave his judgment on the second half of AI: the urgent task is to establish a long-term, AGI-based organization in China.

This organization should be like a balanced triangle. One corner is the foundation model team, making pre-training and post-training solid; one corner is the product team, letting technology create value for people and society; one corner is the Frontier team, exploring new research paradigms.

In an internet giant with tens of thousands of employees and deep barriers between business groups, to build such an organization, the first step is often not looking up at the stars, but first bowing down to tear down the high walls between departments.

Previously, relevant people complained to us that different project departments each had their own AI middle platform, but superiors required projects to cooperate with TEG (Technology Engineering Group). Several teams squeezed around one table, and in the end nobody was happy.

This phenomenon points to the ailment known as big company disease. Tencent has never lacked resources: cloud, games, fintech, social, and office scenarios are all available, and engineers and professional data are piled up like mountains.

But in the traditional bureaucratic assembly line, these assets are scattered across different business groups and reporting lines. Model teams and business teams spend time on organizational friction over who cooperates with whom and how to report results, which obviously cannot grow combat effectiveness oriented toward AGI.

Therefore, a series of organizational cuts after Yao Shunyu joined targeted these blocked joints. Since then, Tencent's AI-related organizational adjustments have all been building the skeleton for such a team structure.

Taking December last year when Yao Shunyu joined Tencent as the starting line, in less than a year afterward, Tencent successively rebuilt pre-training, reinforcement learning, data, and evaluation infrastructure, abolished AI Lab in March, merged the Large Language Model Department and the Multimodal Model Department into the Foundation Model Department in July. By September, the AI Data Department was also brought under it.

It can be seen that Yao Shunyu, following the research logic of the AGI era, connected the underlying layers of Tencent's algorithms, computing infrastructure, and data, allowing these technical forces to operate around the same set of R&D goals.

With such a foundation, the collaboration between business and models has also made considerable progress.

In the traditional delivery pipeline, the model team trains the model behind closed doors, and then the product team comes to find scenarios, with back-and-forth feedback long and somewhat distorted. Tencent's approach is to directly push products into the R&D process upfront. A typical example is Workbuddy, which was the first to connect to and open free trial use of the latest Hy series models.

This internal collaboration mechanism has now begun to spill over to the product side. Buddy One, unveiled at the end of October, connected products such as WorkBuddy, Tencent Docs, Tencent Meeting, Drive, and ima into the same Agent foundation, with accounts, points, and underlying data all connected.

Once authorized, WorkBuddy can retrieve meeting records from Tencent Meeting at any time and can also read ima's knowledge base. By connecting isolated product entrances into the same Agent foundation, the model gets fresher real scenarios, and products get more complete context.

Data shows that since WorkBuddy launched in March, it has completed a cumulative 9 billion task requests, and the enterprise edition covers half of the Fortune China 500 companies.

From business groups fighting separately to coordinated operations around Agents and foundation models, Tencent is working hard to transform itself into a reserve force adapted to AGI evolution.

03 Time waits for no one

While Tencent accelerates organizational and model iteration, the global capital market's attitude toward AI has already changed.

As mentioned earlier, Google suffered a sharp stock price decline after delivering impressive cloud growth. Microsoft, which released its financial report on the same day, escaped disaster, did not trigger panic among AI skeptics, and was basically flat on the first trading day after the report.

Both companies are increasing AI investment and driving cloud business growth, yet market feedback is completely different. The root lies in cash flow and the bearing capacity of core businesses.

In the quarter, Alphabet's capital expenditure reached as high as $51.36 billion, directly pushing free cash flow to -$9.07 billion, negative for the second consecutive quarter. At the same time, its search advertising growth slowed, and the market could not help but begin to worry that massive AI infrastructure spending is consuming cash while the monetization efficiency of the traditional money printer is declining.

By comparison, Microsoft's free cash flow in the quarter remained positive, so the capital market accepted a higher investment ceiling. By this logic, which investments can generate their own cash flow and which are overdrafting cash flow in the long term is an inescapable question for tech companies.

Looking at Tencent with this question in mind, its Q2 2026 financial report shows that the company's single-quarter capital expenditure soared to 52.78 billion yuan, a year-on-year surge of 176%; R&D expenditure was 27.28 billion yuan, up 35% year on year. Investment as aggressive as Google's, together with still-unclear monetization capability, directly emptied free cash flow for the quarter, recording -13.8 billion yuan (after excluding prepayments for computing power procurement, free cash flow was 37.6 billion yuan).

For Tencent, which has always been adept at cash flow management and accustomed to high-margin, asset-light operations, AI is both a ticket to the future and a burden.

Admittedly, Tencent holds three cash cows in gaming, advertising, and fintech, but under new market scrutiny, whether its rich assets can earn it an exemption still depends on third-quarter performance.

Yet at this very moment, the overseas computing power supply chain has produced new variables.

On the very day the official Hy4 version was released, according to media reports, Oracle canceled an AI chip leasing agreement with Tencent valued at about $7 billion. Under the original plan, Tencent had intended to rent about 100,000 advanced AI chips over five years through multiple data centers in Southeast Asia. As of publication, neither side had publicly responded.

Changes in overseas computing power have made Tencent's already-advanced adaptation to domestic computing power even more urgent. Tencent Cloud disclosed that after joint tuning with the Hunyuan team, the single-card throughput of domestic chips was forcibly raised to five times the original just two weeks before the model went online.

When Buddy One, which connects the entire office suite, and more AI products are fully rolled out, the inference computing power consumed every day will only jump exponentially. Tencent not only has to train smarter models, but also has to find every possible way to squeeze down the cost of each call.

Over the past two years, the outside world has always been anxious for Tencent, fearing it would move too slowly and miss the boat. Now, Tencent has finally picked up speed, but the market has changed the exam paper.

Whether it is crunching numbers or doing research, Tencent's small quick steps are still far from being able to slow down.

The copyright of the article's cover image and accompanying images belongs to the copyright owners. If the copyright owner believes that their work should not be browsed by everyone or should not be used free of charge, please contact us in a timely manner, and this platform will correct it immediately.