Summary of Key Points
Tencent's recently released Hy3 official version may not be the largest model with the highest number of parameters or the strongest individual capabilities, but thanks to its engineering expertise (such as cost control and stability) and positive feedback from developers, it has allowed Tencent to re-enter the competition for domestic large-scale AI models. Additionally, by bringing in former OpenAI researcher Yao Shunyu to reorganize the R&D team, Tencent has acquired the necessary methods for building AI models. However, there are still two challenges to overcome: first, whether the model can continuously evolve and remain at the forefront; second, whether the native WeChat agents (Agents) can effectively integrate the platform's entry points, tools, and transaction capabilities, transforming the ecosystem’s advantages into competitive strengths in the AI era.
1. Hy3: Not the Best, but Enough to Get Tencent Back in the Game
Hy3 has parameters of 295 billion and individual metrics (such as a context length of 256K) that are not top-tier, yet it is very popular among developers. It quickly ranked eighth on the OpenRouter platform within three days of its release, and after being integrated with Tencent's WorkBuddy office assistant, the number of calls increased dramatically, with queue times exceeding 50%, even driving Tencent's stock price up by 4.82%.
Why is it so popular? The key lies in its practicality. For example, in coding tasks, although it performs poorly in simple tests compared to DeepSeek, it outperforms its competitors in more complex, real-world-like tests. Enterprises find it cost-effective for high-frequency, medium-to-low difficulty tasks (such as writing documents and creating tables) because it uses the MoE architecture, which activates only 21 billion parameters per inference (like using just some of the “engines”), making it both fast and efficient.
However, the free trial period is still ongoing. The real test will come once charging begins: Will developers be willing to pay for its use? At least, Hy3 has moved Tencent from being excluded from the first tier to a position where developers are willing to give it a try, thus gaining an entry into the AI model competition.
2. Tencent Finally Knows How to Build Models
Over the past two years, Tencent's Hyun Yuan models have lacked prominence: although the parameters and features have been improved, there was never a clear reason to choose them over others. The problem wasn’t a lack of use cases or data, but rather an inability to convert these resources into effective model evolution. Previously, Tencent followed a “big and comprehensive” approach, developing general-purpose models before integrating them with business applications, leading to significant homogenization.
With Yao Shunyu’s involvement, Tencent has changed its strategy: first determining what the products need from models and then figuring out how to train them accordingly. For instance, for office scenarios that require coding skills and tool integration, Hy3’s features are tailored specifically. Feedback from users (such as bottlenecks when using Hy3 for creating tables) is directly incorporated into the training process, continuously improving the model.
This new approach is beginning to show results: not only can Hy3 be used effectively, but internal products (like Yuan Bao and CodeBuddy) are also willing to assign real tasks to it. Tencent has now established a cycle of “building models → using them → optimizing them.”
3. Models Alone Are Not Enough; Two More Challenges Need to Be Overcome
Hy3 is just the first step. For Tencent’s AI to truly succeed, two additional capabilities are needed:
First, can the model continue to evolve? A single successful iteration does not guarantee long-term success. Companies like DeepSeek and Zhipu have been able to release consistently impressive models, so Tencent needs to prove that Hy3 is more than a one-off sensation. This requires sustained collaboration among teams, computing power, talent, and user feedback—something that cannot be achieved with just a single architectural adjustment.
Second, can the native WeChat Agents be successfully implemented? Tencent’s greatest advantage is its WeChat ecosystem, with 1.4 billion users, mini-programs, a payment system, and enterprise processes. However, these elements have not yet been fully integrated. The WeChat agents (such as “Xiao Wei”) are still in the testing phase. Can they help users book flights, use mini-programs to complete tasks, or even make payments directly? If not, Tencent’s ecosystem advantages will be wasted.
For example, a model with a score of 90 that can integrate WeChat payment and mini-programs might be more valuable than one with a score of 100 that can only perform basic chat functions. It would be capable of helping users complete commercially valuable tasks like making purchases or scheduling services.
4. Tencent’s Ace Card: Turning the WeChat Ecosystem into an AI Super Assistant
Compared to other companies, Tencent has a more comprehensive AI strategy: it has models (Hun Yuan), entry points (WeChat), tools (mini-programs), and transaction capabilities (payment). Alibaba has e-commerce and cloud services but lacks a social platform; ByteDance has traffic but a weaker tool ecosystem; DeepSeek has a strong model but no integrated ecosystem.
However, to turn these advantages into a competitive edge, all these components must be seamlessly connected. For instance, if a user asks WeChat to book a train ticket for tomorrow’s trip to Shanghai, the agent should be able to use the 12306 mini-program to check tickets, purchase them with WeChat payment, and even remind the user of the departure time. In this case, the Hun Yuan model acts as the “brain,” while the WeChat ecosystem provides the necessary “arms and legs” to complete the task.
If Tencent can achieve this, its AI will not just be another large-scale model; it will become a “super assistant” that can help users with their daily tasks. This is where its true competitiveness lies.
Conclusion: The Challenges for Tencent’s AI Are Yet to Come
Hy3 has brought Tencent back into the competition, but the road ahead is long. It must prove that the model can continue to evolve and that the WeChat Agents can be effectively implemented. If both goals are achieved, Tencent can transform its ecosystem advantages into a competitive advantage in the AI era. Otherwise, Hy3 might just be an expensive technology investment, and the WeChat ecosystem may not serve as a solid defense. Tencent’s AI journey is just beginning.