Tencent’s Hy4 Delivers a Generational AI Leap as Goldman Reaffirms HK$670 Target
Tencent has delivered its most significant AI model generational leap to date, with the Hunyuan Hy4 preview triggering immediate infrastructure strain and prompting Goldman Sachs to reaffirm a HK$670 price target — implying 47.2% upside from current levels — as the company's tightly integrated product-model strategy begins to distinguish it from peers in an increasingly crowded large language model landscape.
Released on August 28, 2026, Hunyuan Hy4 preview arrived ahead of analyst expectations, sustaining what Goldman Sachs analysts Ronald Keung and Lincoln Kong describe as an approximately two-month cadence since the Hy3 preview. The model's debut was anything but quiet: within hours of launch, Tencent's AI workplace platform WorkBuddy reported task queuing backlogs, forcing the company to execute an emergency expansion of its inference cluster. Even after that scale-up, Tencent acknowledged peak-period queuing may persist — a demand signal that carries its own analytical weight.
Goldman Sachs, in a report dated August 31, characterized Hy4 preview as a "generational capability leap" for the Hunyuan family and identified the model's trajectory as one of the key price catalysts for Tencent's stock over the next 12 months. The bank maintains its sum-of-the-parts (SOTP)-based 12-month target of HK$670, against a current price of HK$455.20.
Parameter Scale Jumps 2.6x, Coding Rank Vaults From 28th to 6th
The architectural step-change in Hy4 preview is quantifiable and substantial. Total parameters expanded from 295 billion (with 21 billion activated) in Hy3 to 770 billion total parameters with 49 billion activated — a roughly 2.6-fold increase. Context window length extended from 256,000 tokens to 1 million tokens, a fourfold expansion that directly addresses enterprise use cases involving large codebases, lengthy financial documents, and complex multi-step agentic workflows.
The most striking benchmark shift is in coding. On the Code Arena WebDev global leaderboard, Hy4 preview ranks sixth globally, compared with Hy3's 28th-place standing — a 22-position jump that repositions Hunyuan firmly within the open-source first tier alongside models such as GLM-5.3 and Kimi K3. In an internal blind evaluation conducted by Tencent using 163 domain experts across 203 real-world work tasks, Hy4 preview scored an average of 2.99, edging out Kimi K3 at 2.94 and GLM-5.3 at 2.92.
Real-world testing corroborates the benchmark data. In structured comparative tests, Hy4 preview produced a six-page equity market review report — richer in content than Hy3's output, incorporating sector performance data, median individual stock returns, and trading volume breakdowns — though it required approximately 44 minutes to complete the task versus Hy3's 10 minutes. In a separate debugging exercise involving a multi-bug data dashboard, Hy4 preview identified all core issues plus additional code maintainability problems, ultimately logging 47 functional checks and 38 cross-device browser tests before delivering a corrected build with test scripts and screenshots. Thoroughness came at a cost: the task consumed roughly one hour of compute time.
Cost Efficiency Holds Despite Scale Expansion, Architecture Innovations Underpin Economics
A critical concern for enterprise customers evaluating large-scale model deployment is whether capability gains come with proportional cost increases. On this metric, Tencent has managed a notable balance. Hy4 preview is priced at approximately US$0.45 per million tokens on a blended basis, competitive with open-source peers of equivalent scale. On WorkBuddy, the model's input and output pricing stands at RMB 6 and RMB 18 per million tokens respectively — six times and 4.5 times the pricing of Hy3 — reflecting the model's expanded capability tier while remaining within enterprise budget parameters for high-value tasks.
Three architectural innovations support this efficiency profile. Tencent introduced a Gated DSA (Differential Sparse Attention) mechanism, drawing on design principles from DeepSeek and GLM, which reduces redundant computation in long-context scenarios. IndexCache enables cross-layer index reuse to lower the computational overhead of processing 1-million-token inputs. iHC (identity Hyper-Connections) improves inter-layer information flow, enhancing reasoning coherence without proportional compute cost. Together, these mechanisms allow Tencent to scale model size while preserving a cost-per-token advantage over comparable open-source alternatives.
"Product-Model Flywheel" Constructs a Data Moat Rivals Cannot Easily Replicate
Goldman Sachs identifies Tencent's most structurally durable competitive advantage not in raw model scale, but in the closed-loop data architecture connecting its model development to its live product ecosystem. The mechanism is straightforward in description but difficult to replicate in practice: Hy4 preview is distributed first through Tencent's own AI-native applications — WorkBuddy for enterprise productivity and CodeBuddy for software development — where real-user interactions at scale generate task trajectories, evaluation signals, and failure modes. That data feeds directly back into subsequent pre-training and post-training iterations, compressing the feedback loop that typically separates model developers from deployment realities.
Tencent's second-quarter 2026 capital expenditure of RMB 52.8 billion (US$7.33 billion) — up 176% year-over-year — provides the financial foundation for this strategy. The company has explicitly stated that a significant portion of that quarter's capex was deployed as AI-related prepayments supporting Hunyuan model upgrades, WorkBuddy, and WeChat AI initiatives.
The training data architecture for Hy4 preview reflects this integration directly. Tencent's Hunyuan team co-developed training datasets with internal domain experts across software engineering, gaming, finance, and security. In an office productivity context, this means training on tasks such as extracting data from unstructured documents, building financial models in spreadsheets, and generating presentation materials — workflows that Tencent employees encounter daily. In gaming, training scenarios include generating playable game prototypes from natural-language specifications and iterating within a game engine. These are not synthetic benchmarks; they are production workflows, and the distinction matters for agentic AI performance in deployment.
Near-Term EPS Pressure Clouds a Structurally Constructive Medium-Term View
Goldman Sachs is explicit about the near-term earnings trade-off. Rising capital expenditure intensity and the long investment horizon of AI infrastructure are expected to compress profit growth in the second half of 2026. The bank projects Tencent's earnings per share growth at just 4% year-over-year in Q3 2026 and 0% in Q4 2026 — a marked deceleration that investors must weigh against the medium-term valuation case.
The bull thesis rests on three pillars beyond Hy4 itself: continued strong user engagement metrics on WorkBuddy, the gradual rollout of Xiaowei — WeChat's AI assistant powered by the separately developed WeLM-80B model — and AI-driven improvements in advertising monetization and gaming.
The existence of WeLM-80B, a WeChat-native model developed independently of Hunyuan, introduces a structural question that Goldman Sachs acknowledges the market has not fully resolved. Tencent President Martin Lau has stated publicly that WeChat AI does not require the most powerful general-purpose model; what matters is fit for WeChat's specific product requirements, including privacy architecture, cost efficiency, and native integration with Mini Programs. The two-model strategy — Hunyuan as the general-capability foundation for WorkBuddy and enterprise cloud, WeLM as the WeChat-optimized layer — reflects a deliberate product philosophy. Whether it represents resource duplication or rational specialization is a question Tencent management is expected to address at the Goldman Sachs Asia Leaders Conference fireside chat on September 1, 2026, where investor focus will center on compute resource allocation, depreciation management, and the scaling roadmap for both WorkBuddy and Xiaowei.
Related Coverage:
Tencent's AI Pivot: What the "Midway Moment" Thesis Really Means