Recently, Tencent officially launched the preview version of HuanYuan Hy4. The release of this version not only came earlier than Morgan Stanley's previous expectation of September, but also directly fulfilled the commitment made by Tencent's management during the second-quarter earnings call to launch it within the year. Since the appointment of Yao Shunyu, Chief AI Scientist, in the fourth quarter of 2025, the update speed of Tencent's model has significantly improved. From the release of the Hy3 preview version in April, the official release of Hy3 in July, to the early appearance of the Hy4 preview version, the market generally expects that the official release of Hy4, originally scheduled for the fourth quarter of this year, will likely be launched ahead of schedule.

In terms of technical architecture, the HuanYuan Hy4 preview version adopts a mixture of experts architecture, achieving a significant leap in model size and long-text processing capabilities. Its total parameter count reaches 770 billion, doubling the 295 billion parameters of the previous generation Hy3; the activated parameter count is increased to 49 billion, higher than the 21 billion of the previous generation; the context window now supports up to 1 million tokens, far exceeding the 256 thousand tokens of the previous generation.

image.png

In terms of performance, with the support of high-quality training data and software-hardware collaborative design, the Hy4 preview version has already secured a position in the industry's top tier. In Tencent's internal double-blind evaluation, the model achieved an average score of 2.99 out of 4, slightly surpassing similar products; in the third-party competition arena's code ability ranking, the model also ranks among the top. In addition to conventional benchmark tests, the model's productivity improvements in real-world scenarios such as software development, office efficiency, game development, and research are also significant. These core advantages are highly aligned with Tencent's main businesses such as gaming, advertising, and cloud services, and are expected to translate into long-term commercial growth.

Based on Tencent's significant progress in the field of artificial intelligence and its solid fundamentals, Morgan Stanley maintains an attractive view of the industry and reiterates its "overweight" rating, setting a target price of 550 HKD. The institution expects that between 2026 and 2028, Tencent's revenue and core profitability will maintain steady growth, with expected price-to-earnings ratios and return on equity demonstrating strong market competitiveness.