Dowson Tong Previews That the Official Version of Hunyuan Hy4 Will Add Visual Understanding for the First Time

Smart Things (Zhidongxi) reported on October 29 that today, at the 2026 Tencent Global Digital Ecosystem Summit, Dowson Tong, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industries Group, delivered a keynote speech, sharing the latest progress on Tencent's AI Agents and infrastructure construction around topics such as the large-scale enterprise deployment of AI Agents, the upgrade of Tencent Cloud's AI products and infrastructure, and the globalization strategy.

Dowson Tong announced that the official version of Tencent's Hunyuan large model Hy4 is about to be released. Compared with the Hy4 Preview released and open-sourced on August 28 this year, the new version has achieved comprehensive improvements, especially strengthening the handling of long-horizon tasks and real productivity tasks, significantly improving chain-of-thought efficiency, and adding visual understanding capabilities for the first time. Tencent will also launch Hy4 Flash in the near future, which, on the basis of inheriting Hy4's visual understanding, Agent and other capabilities, further pursues extreme cost-effectiveness.

On the AI application front, Tencent officially released the cross-application AI collaboration system Buddy One and Tencent Drive.

Buddy One will connect office collaboration products such as Tencent Docs, Tencent Meeting, Tencent Drive, and ima, share the WorkBuddy Agent foundation, and unify accounts and credit pools. More products will join Buddy One later.

English text is being prepared.

At the same time, Tencent Cloud will further connect the desktop office agent WorkBuddy, WeCom, and enterprise internal management systems, enabling Agents to move from assisting in content generation to participating in business analysis, business decision-making, and task execution. Currently, WorkBuddy has cumulatively iterated through 100 major versions and completed 9 billion user task requests.

Speaking about the development direction of enterprise AI, Dowson Tong believes that the key to enterprises using AI is shifting from "adopting it" to "using it well." What enterprises need is not only stronger models, but also Agent systems that can penetrate business processes, continuously create productivity, and improve operational efficiency and return on investment.

In addition, Dowson Tong disclosed that over the past year, the number of overseas active customers of Tencent Cloud's international business has grown about threefold, and overseas AI demand has achieved triple-digit growth. Tencent Cloud will accelerate the layout of global data centers and promote the overseas expansion of AI products such as WorkBuddy, ADP, and Agent Runtime.

The core messages and key quotes of this speech are summarized as follows:

1. Agent explosion: This year is a year of comprehensive explosion in Agent usage. Employees using AI Agents to improve efficiency has already become a mandatory choice.

2. Organizational change: Within Tencent, Agent-based office work has already covered all employees and all scenarios. While cultivating and identifying "super individuals," Tencent is exploring new intelligence-driven organizational methods.

3. Business decision-making: In the AI era, speed itself is competitiveness.

4. Models and cost: Enterprise AI applications are shifting from extensive use of the strongest and most expensive models to precisely matching different models according to tasks and pursuing higher cost-effectiveness. The official version of Tencent Hunyuan Hy4 and Hy4 Flash will be released soon.

The transcript of Dowson Tong's speech is as follows, edited by Smart Things without changing the original meaning:

Good morning, distinguished guests. I am very happy to meet you again at the Tencent Global Digital Ecosystem Summit.

Over the past two years, the development of AI has exceeded almost everyone's expectations, and this year in particular has been a year of comprehensive explosion in Agent usage. According to statistics from the National Data Administration, China's average daily Token calls have grown from 100 billion at the beginning of 2024 to more than 500 trillion in July this year, an increase of more than 5,000 times in two years.

Behind this data is the rapid development of AI applications, and Agents are becoming a new form of productivity.

AI office agents represented by WorkBuddy, in addition to daily document collaboration, are rapidly being deployed across industries in professional fields such as business analysis, business operations, market research, and marketing planning and execution. Employees using AI Agents to improve efficiency has already become a mandatory choice. Enterprises also need AI to make more accurate decisions and execute faster, strengthening their competitiveness in the market; otherwise, they will fall behind if they do not advance.

Since its launch in March this year, WorkBuddy's Token calls have continued to climb. Its enterprise edition has been deployed in more than 50 industries, covering 50% of the Fortune 500 Chinese enterprises and 60% of the top 100 Chinese enterprises. The number of partners and developers on the open platform has also exceeded 50,000, with more than 200,000 Skills listed. At the same time, we have seen that WorkBuddy also performs well in key indicators such as user retention, usage duration, and long-horizon task processing.

I myself am a heavy user of Yuanbao and WorkBuddy. When preparing today's speech, I fed my ideas to WorkBuddy, and it helped me organize materials, extract frameworks, and generate suggestions. We discussed and revised together, achieving "human-machine co-creation."

Within Tencent, using Agents for office work has already covered all employees and all scenarios. WorkBuddy has penetrated deeply into multiple departments and connected internal systems. Colleagues use it every day to complete work, generate and share reports, and have accumulated many professional skills that large models can continuously invoke—and of course, they have also provided the WorkBuddy product team with a lot of feedback and optimization suggestions. We regularly organize cross-department sharing of WorkBuddy use cases and hold internal hackathons to encourage everyone to explore more valuable AI applications. While cultivating and identifying super-individuals, Tencent Cloud is also exploring new ways of organizing.

In the process of building an "intelligence-driven new organization," Tencent Cloud continues to improve its agent product matrix, both meeting the needs of different business scenarios and providing a low-cost, high-performance, secure and controllable multi-Agent runtime environment, creating an Agent productivity system that is "full-scenario, high-efficiency, and more reassuring."

Next, I will focus on these three keywords and share Tencent's exploration and practice in detail.

1. Connect office tools with business systems so that Agents understand business context and integrate into workflows

The first word is "full-scenario."

Enterprise customers often report: the models are already strong and there are plenty of tools, but when actually applied to business, there are still many breakpoints. For example, AI can do analysis and write proposals, but if it cannot connect to business systems to obtain data, it cannot understand the actual situation of the enterprise's business, channels, and customers.

To bridge these breakpoints, connection must come first. WorkBuddy connects people, tools, systems, and data, enabling large models to understand business context and integrate into actual workflows.

1. Launch Buddy One and Tencent Drive to unify access to office tools and save work outputs

First is connecting office scenarios so that tools collaborate around tasks.

Over the past few decades, digital office work has mainly used files as the carrier. Different types of file formats are browsed and edited by different applications, all designed for people using a mouse and keyboard input. There are many desktop tools, documents are too scattered, and editing takes a relatively long time. During multi-person collaboration, each person modifies their own version and then circulates it back and forth, which is inefficient and quite troublesome.

But in the era of office Agents, I believe multi-person collaboration will enter a new paradigm of human-machine co-creation: after the team discussion, requirements are submitted to AI, content is generated by large models, integrated by Agents, accepted by the team, and then stored in the knowledge base.

Today, taking the opportunity of the ecosystem conference, we are also officially launching Buddy One externally.

Buddy One uses WorkBuddy as the workbench and seamlessly integrates different tools such as Tencent Docs, Tencent Lexiang, and Tencent Design. Through a unified account, unified credit pool, and unified data pool, large models invoke different tools for different needs to generate multimodal content. Teams can collaboratively modify within projects, jointly inspect outputs, and continuously discuss and iterate. Buddy One enables human-AI co-creation, collaborating around tasks rather than around file circulation. Of course, it also needs to support linkage across many terminals, with seamless switching of Agents between cloud and device.

Such an "integrated experience" requires the support of multiple products. Tencent Docs handles content editing, Tencent Lexiang provides knowledge management, and Tencent Meeting incorporates the consensus from meetings into to-dos—all providing context for WorkBuddy.

When a task needs supplementary external information, Yuanbao Search also provides real-time data. Relying on authoritative domestic and international websites, professional data resources, and Tencent's unique content ecosystem, such as Tencent Medipedia, Yuanbao Search has accumulated an index of hundreds of billions of entries to meet various deep research needs. Yuanbao not only serves as a ChatBot providing users with fast and accurate answers, but also provides search capabilities to the industry through APIs.

English text is being prepared.

2. WorkBuddy connects to WeCom and business systems, supporting intelligent data querying

Second is connecting data platforms so that Agents enter enterprise workflows.

In the past, various enterprise management systems such as ERP and CRM accumulated a large amount of enterprise operational data, with specific UI interfaces for viewing data and managing processes. But each system is cut off from the others; although each has some API interfaces, cross-system data analysis—even if the systems are connected—involves enormous workload and long waiting times each time.

Today, with WorkBuddy, by having these management systems open MCP interfaces, as long as the corresponding skills are established and the data structures and relationships are clearly described, the Agent can break down various business analysis tasks, proactively obtain the needed data from multiple systems, automatically generate the required reports, and present the analysis results. Of course, we can also build different task Agents to asynchronously submit data to systems—for example, organizing invoices in batches and submitting reimbursements in batches—without the need to synchronously operate UI interfaces as before, which is slow and cumbersome.

There are also many enterprises where employees discuss project division of labor in WeCom and reach consensus—this is very important contextual input for office Agents.

This month, WorkBuddy and WeCom jointly released a new version, achieving the integration and upgrade of AI office work and mobile office work. WorkBuddy has officially entered the WeCom sidebar as an intelligent robot, and employees can hand tasks to it at any time, just as they would talk to a colleague. WorkBuddy can work 24 hours a day and send the results directly back to WeCom when done. Employees can also log in to WorkBuddy with their WeCom accounts and authorize it to read data and execute tasks.

As shown in the DEMO, by issuing a simple instruction through the WorkBuddy APP, it can, within the scope of authorization, check unread messages and emails, compile key matters and to-dos, and push them to the phone; when needed, it can also generate reply suggestions and send them with one click after confirmation. For employees who travel frequently and have dense meeting schedules, they can efficiently handle work in fragmented time.

After WorkBuddy is connected with WeCom, users' communications, documents and meeting records in WeCom can also, within the scope of authorization, become the basis for the Agent to understand tasks and carry out work. Enterprises do not need to change systems, migrate data or redo processes; they can directly turn digital assets accumulated over many years into AI production materials. Managers can also uniformly allocate usage seats, Token quotas and access permissions, and reclaim them uniformly after employees leave.

At the same time, WorkBuddy will continue to open up its product architecture, supporting enterprises' self-developed Agents and third-party Agents in the ecosystem, including access to third-party IM systems. Which model to choose, which Agent to connect and which IM to use—the choice is in the hands of enterprises. More and more enterprises are building their own enterprise Agent networks based on WorkBuddy and constructing a new generation of enterprise intelligent hub.

As just mentioned, for an Agent to master more enterprise data and context, it must also further connect to enterprises' internal management systems, such as ERP, CRM, OA and financial software, among others.

Take Tencent Cloud as an example. We have built an operations service intelligent agent, "Panshi Xiao O." In the past, when our colleagues wanted to check a customer's operating data, they had to find the system, go through approval and dig through reports; now, by asking Xiao O a question in WorkBuddy, they can, within the scope of permissions, check sales performance, customer operations and collection progress. Relying on WorkBuddy's "hosted agent service" (WMA), Xiao O runs through unified identity authentication, permission control and a 7×24-hour isolated environment, making every operation of the Agent auditable and traceable and ensuring the security of operating data.

Intelligent data querying is very valuable for management decision-making. I have personal experience of this. For example, in the past, it took a data engineer two to three weeks to produce a report; after managers saw it and wanted to follow up on business details and look for optimization directions, they had to wait for the report to be updated again, making it easy to miss decision-making opportunities. If a decision was made hastily, it was hard to guarantee its quality. Now, AI-native managers can ask questions through Agents at any time, keep abreast of business dynamics in a timely manner, and see problems, make judgments and take action faster. For managers, this kind of decision-making efficiency improvement is key for enterprises to maintain long-term competitiveness. Those who can better use AI to make faster decisions and act earlier may discover business opportunities earlier in the market or gain more share in the existing market. In the AI era, speed itself is competitiveness.

For every employee, after the Agent is connected with operating systems, it can not only query data but also handle more matters. For example, before a salesperson visits a customer, Xiao O can, based on CRM data and public information, compile customer dynamics and communication suggestions; schedule a meeting and sync the calendar with one sentence; and after the visit, aggregate business data and generate reports and to-dos. When approval is needed, it can also initiate it directly.

3. MAGIC connects marketing and ordering, ADP supports customized service Agents

Finally, there is building business Agents to execute operational tasks, connect partners and serve customers.

From internal collaboration to business operations, Agents can also continuously execute tasks against clear goals and quantifiable indicators. For example, in scenarios such as customer service, marketing and transactions, enterprises need professional business Agents: they follow business rules and processes, operate around the clock, and not only complete tasks but also serve operators and be responsible for actual business results.

Recently, the Tencent Qidian Marketing Cloud MAGIC Agent has launched a Buddy application on the WorkBuddy open platform. Taking promotions of popular products as an example, WorkBuddy compiles the product list, MAGIC connects to the customer data platform CDP, selects audiences and executes marketing activities, and Cloud Mall handles orders; a single entry point can string together the entire process of preparation, execution, transaction and review.

Enterprises can also use Tencent Cloud ADP to customize more personalized service Agents and publish them to the enterprise Agent unified portal AgentPortal, or directly embed them into enterprise official websites, mini programs and business systems. We are also further promoting deep integration between ADP and WorkBuddy, so that more business Agents can connect to WorkBuddy and become capabilities that employees can invoke at their fingertips.

At Yili, they are combining ADP, WorkBuddy and WeCom for marketing and sales: WorkBuddy aggregates business data, analyzes it and generates operating strategies, which are distributed to thousands of shopping guides, improving operational efficiency by more than 50%; the ADP shopping guide Agent can generate product recommendations and communication scripts in real time, and A/B testing shows a 26% increase in sales orders.

2. Hunyuan Hy4 will add visual understanding, Flash will focus on cost-effectiveness, and Agent Runtime will be upgraded simultaneously.

Next, let's talk about the second word: "high efficiency."

As Agents continue to penetrate office and business scenarios and begin to support enterprise-level production, enterprises' token usage and costs have surged, creating an urgent need for more refined token resource management to reduce costs. The shift is from extensive use of the strongest and most expensive models to pursuing cost-effectiveness and precisely matching different models to execute different tasks.

Tencent Cloud has high-performance and extremely cost-effective Agent infrastructure, stable heterogeneous computing clusters, and deep cooperation with multiple chip manufacturers to build a new ecosystem of domestic computing power.

Amid global computing power constraints, Tencent Cloud has deeply adapted to chips from multiple manufacturers, continuously optimizing inference performance on domestic chips to achieve excellent cost-effectiveness, and has been widely used in manufacturing industries such as embodied intelligence. To handle the peak call volume after the release of Hy4 Preview, Tencent Cloud worked closely with the Hunyuan team and, in the two weeks before launch, increased the single-card throughput of a certain domestic chip by 5 times; we also helped a leading international Tier 1 automotive manufacturer deploy a domestic thousand-card cluster to support large-scale training of end-to-end intelligent driving large models.

On the model side, Tencent Hunyuan Hy continues to evolve in real business scenarios. Over the past year, Hy rebuilt its data and training infrastructure and released a major version on average every two months. Hy4 preview is positioned as "born for productivity," and is very stable and reliable on real productivity tasks such as coding, office work, and science. Since its release, it has consistently ranked among the top on OpenRouter and has supported the stable service of products such as Yuanbao and WorkBuddy.

The Hunyuan large model Hy4 will be released soon. Compared with Hy4 preview, it achieves comprehensive and substantial improvements, especially strengthening capabilities for various long-horizon and real productivity tasks, significantly improving chain-of-thought efficiency, and introducing visual understanding for the first time.

Hy4 flash will also be released soon, providing a practical model with ultimate cost-effectiveness. On the basis of inheriting Hy4's strong capabilities such as vision and Agent, it will evolve at high frequency through deep Co-Design of models and products.

Tencent always adheres to an open strategy. Based on the TokenHub platform, users can call mainstream open-source large models such as Hunyuan, DeepSeek, GLM, MiniMax, and Kimi through APIs. With the unified billing service of Token Plan, enterprises can choose among multiple models on demand at extremely low switching costs.

In content production scenarios, we have also launched the Tencent Cloud audio-video AI WAND product family, providing mainstream multimodal models, the WAND-Vega multimodal application model, and more than 100 self-developed media processing tools, covering 90% of leading short comic drama overseas customers, generating 4.5 million images and 3 million seconds of video per day on average.

In addition, in terms of Agent Infra, today we have also fully upgraded Agent Runtime, making systematic enhancements around key areas such as context management, tool invocation, secure sandboxing, observability, and evaluation, so that Agents run more stably, more controllably, and with higher overall cost-effectiveness. My colleague will provide a detailed introduction later.

3. Intercept high-risk operations, record invocation processes, and support encryption and recovery of enterprise data.

The last keyword is "more peace of mind."

As more and more Agents are deployed within office networks, both Agents and large models will come into contact with more enterprise-level data, making it particularly important to ensure the security of networks, runtime environments, and data. Enterprises must not only prevent sensitive data from leaking to third parties, but also prevent malicious attacks from internal and external users, as well as protect intellectual property and prevent data theft beyond authorized scope.

AI also amplifies the capabilities of bad actors, and Agents greatly lower the threshold for wrongdoing. Recently, we have seen a surge in the number of enterprise root keys being stolen, with affected organizations later coming to our security team to fix the problems and deploy a new generation of security products to strengthen protection.

Tencent itself operates many national-level basic services. We know that security is the foundation of trust, so we always put security first in our products and services. For Agent scenarios, we are committed to building a "trusted AI" governance system, mainly covering three aspects.

First, behavior is controllable. Who the Agent represents, what it can see, and what it can do must be clearly defined first; during operation, it must be monitorable, intervenable, and stoppable. In WorkBuddy, hidden malicious instructions are identified and cleaned, extension capabilities must pass security verification before loading, high-risk operations such as data deletion and unauthorized access are directly intercepted, and uncertain operations are left to users for confirmation.

Second, processes are traceable. Each Agent has an independent identity and is incorporated into the enterprise's unified account and permission system; what data it has viewed, what tools it has invoked, and what actions it has executed are all recorded throughout the process. Administrators can locate problems by person, by time, and by Agent, and can also uniformly configure rules to restrict data outbound transmission.

Third, data is manageable. What enterprises worry about most is data: where it is stored, how it is transmitted, and what to do if it is accidentally deleted. Our answer is full-process encryption and fallback recovery: private data is encrypted throughout the process from endpoint to transmission to storage, and sensitive tasks support confidential computing; in the event of accidental deletion, overwriting, or damage, file-level recovery can be performed, so that a single accident does not become an irreversible loss.

These capabilities have also received broad recognition from the outside world. WorkBuddy is among the first batch of products in China to pass the CAICT "Trustworthy AI" assessment, and it also received the CCIA (China Cybersecurity Industry Alliance) "Highest-Level Personal Information Protection Evaluation." At the same time, as a core unit, we also participated in the compilation of Shenzhen's "Standard System for Trustworthy AI Office Agents," providing best practices for trustworthy AI office work for standard-setting. These security capability-building efforts have also won recognition from government and enterprise customers such as the Guangdong Provincial Government Services and Data Management Bureau, Shenzhen Airlines, and Shenzhen Tong.

IV. Overseas AI demand grows by triple digits, accelerating data center layout and Agent product globalization

Tencent is also actively bringing the domestic experience of deploying large models and AI Agents to the world, from the computing foundation and model services to Agent-driven technology platforms and industry solutions, helping more global enterprises build cost-effective, secure, and controllable AI technology stacks.

Over the past year, the number of overseas active customers of Tencent Cloud International has doubled, covering more than 150 countries and regions worldwide. In several key markets, we have achieved depth in our business:

In Indonesia, we have successively signed and are actively advancing comprehensive cloud and AI cooperation with the three major local operators XLSMART, Telkomsel, and Indosat.

In Thailand, we continue to help more business segments of the Charoen Pokphand Group reduce costs and increase efficiency, accelerate intelligent transformation, and have entered leading operators, exchanges, and banks.

In Hong Kong and Macao, we have served customers such as Hung Fook Tong and the Government Digital Policy Office, with business maintaining steady high-level growth for seven consecutive years.

In Europe, we continue to expand deep cooperation with global internet platform companies such as iyzico and eMAG.

First, we will accelerate the construction of global data centers.

Tencent Cloud's overseas region-opening speed has long remained industry-leading, with more than 2 million servers deployed globally. Over the past year, we have added or expanded data centers in Frankfurt, Malaysia, Mexico, Osaka, Riyadh, and Thailand. In the coming year, we will further expand our layout in Asia, the Middle East, Europe, and the Americas.

Second, build cost-effective infrastructure.

Facing the surge in data, general computing, and storage demand brought by the AI wave, we continue to design optimal software and hardware architectures for customers. For example, the latest flagship Star Lake whole machine supports 1024 cores, the cloud supports upgrading cloud servers to the latest SA10, and with extreme memory management technology, we provide customers with a highly cost-effective computing platform.

Facing rising storage costs, we have also designed tiered storage solutions for many customers; currently, COS tape storage has accumulated more than 2,000 PB, helping global customers archive "extremely cold data" at low cost.

For customers who have their own data centers and want to upgrade from virtualization technology to a cloud-native foundation, the minimum deployment scale of the dedicated cloud TCE has been reduced from 14 units to 8 units, further lowering the deployment threshold and cost.

At the model level, domestic open-source models have demonstrated a good experience and extremely high overall cost-effectiveness in high-frequency scenarios such as office work, knowledge Q&A, and content generation. After Hy4 Preview was launched, it quickly received inquiries from many overseas enterprises. In the future, we will also invest more resources to ensure the global supply of high-quality Tokens.

At the product level, we are accelerating the globalization of various Agents such as CodeBuddy, WorkBuddy, Tencent Design Miora, ADP, MAGIC AI, and Agent Runtime. Whether it is the overseas-assigned employees of Chinese enterprises or the local teams of overseas customers, they can all collaborate on the same AI workbench.

In serving overseas customers, we continue to assist overseas platform enterprises in building local Superapp ecosystems. The upgraded mini program platform solution Superapp as a Service has achieved deep synergy with CodeBuddy, allowing developers to generate mini programs using natural language. Indonesia's local operator Telkomsel has built a mini program ecosystem based on it, with MAU doubling within a year.

In audio and video scenarios, during this year's FIFA World Cup in the United States, Canada, and Mexico, WAND audio and video AI technology provided technical services to official broadcast platforms in 17 countries and regions. During the Nagoya Asian Games, the President of the International Olympic Committee praised on site the Chinese broadcasting technology we co-created with China Media Group.

In cloud migration scenarios, we have developed more than 20 migration Skills based on CodeBuddy and WorkBuddy. In only 4.5 months, we helped Indonesia's leading operator XLSMART complete a smooth migration, while also developing more than 200 localized features.

Today, the global competitiveness of Chinese technology enterprises no longer comes only from cost, but more from the comprehensive capabilities of AI, cloud, and engineering delivery. Tencent Cloud hopes to become a bridge connecting China's AI with global industries, helping more enterprises achieve growth in the global market.

Distinguished guests and partners, looking back at the development of information technology, real industrial transformation comes not only from technological breakthroughs, but also from technology going deep into business, strengthening connections with upstream and downstream, and even changing the way organizations work.

The same is true for AI. From "using it" to "using it well," the key is to let AI enter processes, create productivity, and continuously improve operational efficiency and return on investment. Tencent will adhere to long-term investment, continuously refine products, open up platform capabilities, and help enterprises build operable and evolvable Agent productivity systems.

In this process, we also always keep our doors open to partners and look forward to working with everyone to jointly build a prosperous Agent industry ecosystem.

Thank you all.