Goldman Sachs is bullish on Hy4: Significantly improved coding and agent capabilities; "product + model closed loop" becomes the key to Tencent AI differentiation.

Goldman Sachs believes that the release of the Hy4 preview version marks the largest generational leap in the model family to date, and regards Tencent's unique "product + model closed loop" strategy as the core path for it to establish a differentiated competitive advantage in the AI Agent era.

According to a research report released by Goldman Sachs on August 31, the release of the Hy4 preview version was earlier than analysts expected, continuing the release schedule of approximately once every two months since the Hy3 preview version. In terms of coding capabilities, the Hy4 preview version jumped to 6th place on the Code Arena WebDev global leaderboard, while Hy3 only ranked 28th, demonstrating a particularly significant improvement.

Goldman Sachs analysts Ronald Keung, Lincoln Kong, and others believe that this development will be one of the key drivers of Tencent's stock price over the next 12 months, and maintain their 12-month target price of HK$670 based on the SOTP method, implying an upside of approximately 47.2% from the current share price of HK$455.20.

Despite rising capital expenditure intensity and long-term AI investment suppressing near-term profit growth—the bank expects Tencent's earnings per share to grow by only 4% and 0% year-on-year in the third and fourth quarters of 2026, respectively—the continued iteration of the Hunyuan model, the strong performance of WorkBuddy user metrics, and the gradual promotion of WeChat Mini Programs will jointly support Tencent's medium-term valuation logic.

The parameter scale has been greatly expanded, and the coding capability has achieved a generational leap.

The Hy4 preview version of Hunyuan has achieved a significant expansion in model size, with the total number of parameters increasing from 295 billion/21 billion activation parameters in Hy3 to 770 billion total parameters/49 billion activation parameters, an increase of about 2.6 times; the context length has also been expanded from 256K tokens to 1 million tokens.

Tencent characterizes this release as the most significant generational leap in the history of the Hy4 model family, with substantial expansion in both pre-training and post-training. In terms of coding capabilities, the Hy4 preview version ranks 6th on the Code Arena WebDev global leaderboard, a significant jump from Hy3's 28th place, propelling Hy4 back into the top tier of open-source models.

In terms of professional task coverage, the Hy4 preview version has achieved meaningful capability enhancements in multiple fields, including software engineering, office productivity, game development, and scientific research. Tencent collaborated with internal expert teams to build training data around real workflows using proprietary domain knowledge, and co-designed with workspace products such as WorkBuddy and CodeBuddy to obtain organic user feedback for model iteration.

Cost efficiency maintains a competitive advantage, while architectural innovation supports inference efficiency.

Despite the significant increase in model size, the Hy4 preview version remains competitive in terms of cost efficiency, with a hybrid pricing of approximately $0.45 per million tokens, which is low compared to similar open-source products of the same size, while offering comparable or even better performance.

At the architectural level, Tencent has introduced several inference efficiency innovations, including a gated DSA attention mechanism inspired by DeepSeek and GLM, IndexCache for improving the efficiency of long-context computation, and iHC (identity hyper-connections) for enhancing inter-layer information flow. These architectural innovations help Tencent maintain its cost advantage over similar open-source models while continuously expanding the scale of its models.

"Product + Model Closed Loop" Builds a Differentiated Moat

The report emphasizes Tencent's differentiated strategy: the preview model is distributed first through Tencent's own ecosystem and AI native applications (WorkBuddy, CodeBuddy), data is collected in large-scale real user scenarios, and this data is then fed back into subsequent pre-training and post-training iterations to form a positive data flywheel.

This closed-loop approach is particularly valuable in productivity and coding workload scenarios because real-world task trajectories, user interactions, and evaluation signals are the core drivers of model differentiation in the AI Agent era. Tencent Cloud and Smart Industry Group management recently reiterated in media interviews that Tencent's AI strategy increasingly focuses on creating long-term value through continuous investment in products, scenarios, and user adoption, aiming to solve real-world problems and build lasting differentiation.

Tencent will participate in a fireside chat at Goldman Sachs Asia Leadership Conference on September 1. Investors are expected to focus on how to balance computing power resource allocation with depreciation pressure, the differentiated model strategy of Hunyuan, the scaling path of WorkBuddy, the promotion progress of WeChat Mini Programs, and the prospects of games and AI-driven advertising.

~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Zhuifeng Trading Platform .

For more detailed analysis, including real-time updates and firsthand research, please join the 【 Trader's Trading Platform Annual Membership 】.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.