CITIC Securities International: Tencent Holdings' Mixed Yuan Hy3 enters the first tier as AI product ecosystem accelerates
CITIC Securities International released a research report, maintaining Tencent's "Buy" rating and target price of HKD 632. The firm believes that the official release of Tencent's mixed Yuan Hy3 large model validates its execution capability, and with low computing power costs and high efficiency, it is expected to reduce Tencent Cloud's AI marginal costs and enhance product experience. The report points out that Tencent's AI investment is shifting from expense spending to internal productivity tools, and the market will reassess the implied value of its AI assets
According to the Zhitong Finance APP, China Merchants Securities International released a research report stating that Tencent Holdings (00700) officially launched the Hongyuan Hy3 large model. This model is the official iterative version of the Hy3 preview released in April, continuing the MoE architecture with a total of 295 billion parameters and 21 billion active parameters, along with a context of 256K. The firm maintains an "Overweight" rating on Tencent with a target price of HKD 632. The official release of Hy3 validates Tencent's execution capability in enhancing the reliability of post-training driven models, and the market will reassess the implied value of Tencent's AI assets.
The report indicates that on the cost side, the 21 billion active parameters allow Hy3 to operate at a computational cost far lower than flagship models of similar capability, with the API input price dropping to 1 RMB. WorkBuddy data shows that the actual token consumption in business is about 47% to 49% lower than GLM-5.2, which directly reduces the marginal cost of Tencent Cloud AI services and provides controllable scaling conditions for large-scale applications aimed at consumers, such as Yuanbao, ima, and WeChat Reading.
On the revenue side, the significant enhancement in model capability is expected to drive an increase in API call volume for Tencent Cloud's large models. Additionally, metrics such as a 37.6% increase in the success rate of programming tasks in QQ Browser and a 14.1% improvement in tag accuracy in WeChat Reading reflect the positive impact of the model on specific product experiences, indirectly supporting user engagement and commercialization conversion.
On the ecosystem side, Hy3 deployment has covered core scenarios such as office, Agent, gaming, customer service, and reading. From the end of January to July, the full chain from infrastructure reconstruction to product-level models was completed, indicating that Tencent's AI investment is shifting from expense spending to internal productivity tools, and the boost to R&D efficiency and product iteration speed will gradually manifest in the following quarters
