Goldman Sachs views the release of the Hunyuan Hy4 preview as the most significant generational leap for the model family to date, identifying Tencent's unique "product-model closed loop" strategy as the central path to establishing a competitive edge in the AI agent era.
According to trading desk data, Goldman Sachs noted in a research report dated August 31st that the Hy4 preview launch came earlier than analysts anticipated, maintaining a roughly two-month release cadence since the Hy3 preview. In coding proficiency, the Hy4 preview climbed to 6th place globally on the Code Arena WebDev leaderboard, a notable jump from Hy3's 28th position, underscoring the magnitude of improvement.
Analysts including Ronald Keung and Lincoln Kong believe this progress will serve as a key driver for Tencent's stock over the next twelve months. They reaffirmed a 12-month SOTP-based target price of HKD 670, implying roughly 47.2% upside from the current share price of HKD 455.20.
While rising capital expenditure intensity and long-term AI investments are expected to pressure near-term profit growth—with the bank forecasting year-on-year EPS growth of just 4% and 0% for Q3 and Q4 2026, respectively—the continuous iteration of the Hunyuan model, robust user metrics for WorkBuddy, and the gradual rollout of WeChat Xiaowei are seen as supporting Tencent's medium-term valuation thesis.
Where the Scale-Up Matters Most
The Hunyuan Hy4 preview marks a substantial expansion in model size, with total parameters growing from 295 billion (with 21 billion activated) in Hy3 to 770 billion total (with 49 billion activated)—an approximately 2.6-fold increase. Context length has also been extended from 256K tokens to 1 million tokens.
Tencent has characterized this release as the most sweeping generational capability enhancement in the Hunyuan family's history, with both pre-training and post-training experiencing significant scale expansion. In coding tasks, the Hy4 preview ranks 6th globally on Code Arena WebDev, a sharp rise from Hy3's 28th place, re-establishing Hunyuan within the first tier of open-source models.
On professional task coverage, the Hy4 preview delivers meaningful improvements across software engineering, office productivity, game development, and scientific research. Tencent has collaborated with internal expert teams to construct training data around real-world workflows, using proprietary domain knowledge, while co-designing with workspace products like WorkBuddy and CodeBuddy to capture organic user feedback for iterative model refinement.
Cost Efficiency and Architectural Innovation as Competitive Levers
Despite the larger model footprint, the Hy4 preview maintains competitive cost efficiency, with blended pricing around $0.45 per million tokens—on the lower end among open-source peers of comparable scale—while delivering comparable or superior performance. On the architectural front, Tencent has introduced several inference efficiency innovations, including a gated DSA attention mechanism inspired by DeepSeek and GLM, IndexCache for improved long-context computational efficiency, and identity Hyper-Connections (iHC) to enhance inter-layer information flow. These architectural choices help Tencent sustain cost advantages over similar open-source models as it continues to scale.
A Differentiated Moat Through the Product-Model Loop
The report emphasizes Tencent's differentiated approach—preview models are first distributed through its own ecosystem and AI-native applications like WorkBuddy and CodeBuddy, collecting data across large-scale real-world user scenarios, which is then fed back into subsequent pre-training and post-training cycles. This creates a positive data flywheel.
This closed-loop methodology proves especially valuable in productivity and coding workloads, where authentic task trajectories, user interactions, and evaluation signals are the core drivers of differentiation in the AI agent era. Management from Tencent Cloud and Smart Industries Group recently reiterated in media interviews that the company's AI strategy increasingly focuses on creating long-term value through sustained investment in products, scenarios, and user adoption, aiming to solve real-world problems and build durable differentiation.
What Investors Will Watch Next
Tencent is scheduled to participate in a fireside chat at the Goldman Sachs Asia Leaders Conference on September 1st. Investor attention is expected to center on the balance between compute resource allocation and depreciation pressure, the differentiated model strategy for Hunyuan, the scaling pathway for WorkBuddy, the rollout progress of WeChat Xiaowei, and the outlook for gaming and AI-driven advertising.
Comments