Summary of Key Points
Tencent has released the official version of its Hy3 large model, which marks the first significant achievement since the reorganization of its AI team (led by Yao Shunyu). This represents a shift from a situation where Tencent's models lacked strength and were disconnected from actual products to one where these models can be effectively integrated into business operations and produce tangible results. However, Hy3 still has shortcomings compared to leading competitors, such as the absence of native multimodal capabilities. The ultimate success of Tencent's AI efforts will depend on its ability to transform model capabilities into user experiences and business growth through its various services, including WeChat, gaming, and cloud computing.
I. Hy3: Tencent AI Finally Shows Concrete Results
Hy3 is not just a model that appears on lists of top performers; it represents the first time Tencent has truly integrated model capabilities into its products. For example:
- Office scenarios: The use of Hy3 in WorkBuddy increased task success rates from 72% to 90%, and tasks such as processing 100-page technical manuals, generating test questions, and creating HTML games were completed with a single successful code run.
- Consumer products: After integrating Hy3 into Yuanbao, the error rate of common knowledge claims and the occurrence of hallucinations were both reduced by half.
- Internal data: Since the preview version of Hy3 was launched, the daily volume of text processed (in terms of tokens consumed) increased by 20 times, and the number of users who voluntarily chose Hy3 in WorkBuddy increased by sixfold. In short, Hy3 has transformed Tencent AI from being invisible to solving real problems.
II. Tencent AI’s Path to Improvement: From a “Leaking Ship” to Yao Shunyu’s Leadership
Tencent AI previously faced significant challenges, with its general large models not generating much attention and users perceiving them as underperforming. Ma Huateng even described it as being like a “leaking ship” on which one could stand but not sit comfortably—indicating a lack of foundational capabilities in the early stages.
To address these issues, Tencent took two key actions:
1. Reorganize the team: At the end of last year, departments for AI infrastructure and data were established, with Yao Shunyu (Chief AI Scientist) overseeing both the large language models division. In March this year, the AI Lab was disbanded, and some of its members joined the Hunyuan team.
2. Promote collaboration between technology and product teams: Yao Shunyu instructed the post-training specialists to first help optimize Yuanbao, demonstrating that the tech team was focused on product needs, which laid the foundation for the successful implementation of Hy3.
Now, with Yao Shunyu taking the lead, it signals that Tencent AI is serious about building a solid foundation for its models.
III. Hy3’s Strengths: Practicality First, Solving Real Problems
Yao Shunyu emphasizes that practicality is more important than just appearing on lists of top models. The design of Hy3 focuses on the following aspects:
- Long-text processing: It can quickly understand 100-page technical manuals and extract key information accurately.
- Agent capabilities: In WorkBuddy, it can break down complex tasks into steps and complete them autonomously.
- Product integration: When integrated with WeChat’s customer service, the accuracy of intent recognition reached 98.94%. The WeGame game assistant saw a 92% increase in successful multi-round reasoning attempts, and the hallucination rate decreased from 4.5% to 2.8%.
These improvements are real, practical enhancements, not just impressive numbers from laboratories.
IV. Hy3’s Weaknesses: Still Falling Short of Leading Competitors, Lacking Multimodal Capabilities
Despite progress, Hy3 still lags behind leading competitors:
- Performance: In expert blind tests, Hy3 scored 2.67 out of 4, which is higher than Zhipu’s previous generation GLM-5.1. However, Zhipu’s latest GLM-5.2 offers a larger context window and stronger programming and multi-tool integration capabilities.
- Multimodal limitations: Hy3 can only process text; it cannot directly understand images, videos, or screenshots. For example, if a user sends a PPT screenshot to the AI, Hy3 needs to use OCR tools to convert it into text before processing it. Competitors like DouBao 2.1 Pro and Alibaba Qwen3.7-Plus can already process visual information directly and even support multi-agent collaboration in 3D virtual environments.
These shortcomings mean that Hy3 may not be competitive in consumer-facing scenarios (such as WeChat) or complex fields like chip design.
V. The Next Phase for Tencent AI: Beyond Models, the Application Layer is Key
Hy3 is just a starting point. Whether Tencent AI can truly succeed depends on how well its models are integrated into various applications:
1. Integration with core businesses: Can Tencent’s core services (WeChat, gaming, advertising) leverage Hy3 to create more valuable features for users? For instance, can WeChat develop smarter chat assistants or games with more player-aware NPCs?
2. Competition in the cloud market: While Huoshan Engine’s DouBao has over 180 trillion daily calls and holds a 49.5% share of the MaaS market, Tencent Cloud’s AI cloud market share is only in single digits. Hy3 needs to help Tencent Cloud regain its position in this market.
3. Exploring new consumer scenarios: Yao Shunyu mentions the need for more personalized experiences. For example, answering questions about what to eat today should differ from last year based on user data (such as preferences and location). Although Tencent has this data, it hasn’t yet created AI features that users cannot live without.
In summary, Hy3 has helped Tencent AI take a step forward, but to truly succeed, all business teams must transform model capabilities into tangible value for users. This is the biggest challenge ahead for Tencent AI.
(The translation maintains the original Markdown structure and uses clear, financial journalism-style language that is easy for non-experts to understand.)