Goldman Sachs: Tencent's Hunyuan Hy4 Achieves Generational Leap, Maintains Target Price of HK$670
nashnova research
Goldman Sachs said on August 31 that Tencent's Hunyuan Hy4 preview delivered the model family's largest generational capability jump, maintaining a HK$670 target price — roughly 47% upside — built on Tencent's unique "product + model closed loop" as its core AI-agent-era edge.
What exactly did Hunyuan Hy4 upgrade?
Total parameters jumped from Hy3's 295 billion to 770 billion; active parameters rose from 21 billion to 49 billion — roughly a 2.6× increase.
The context window — how much content the model can process in one pass — expanded from 256K tokens to 1 million tokens.
On the Code Arena WebDev global leaderboard, Hy4 climbed to 6th place; Hy3 sat at 28th. This means → Hunyuan is back in the top tier of open-source models, no longer stuck in the awkward spot of "big-tech in-house but off the charts."
The model got bigger — how are costs kept in check?
Hy4 preview's blended pricing is roughly $0.45 per million tokens, low among open-source products of comparable scale.
On the architecture side, Tencent introduced gated DSA attention — a design that lets the model allocate compute resources more efficiently — along with IndexCache and iHC inference optimizations.
In plain terms = parameters more than doubled, but per-call pricing did not rise proportionally. The savings come from built-in architectural efficiency.
What does the "product + model closed loop" mean?
Tencent's playbook: ship the preview model first through its own AI-native apps — WorkBuddy and CodeBuddy — and get real users on it.
Task traces, interaction data, and evaluation signals generated by those users are then fed back into the next round of pre-training and post-training. This means → more users, faster iteration — Goldman calls it a "positive data flywheel."
This reflects a fundamentally different AI path from pure benchmark racing: Tencent is betting on exclusive access to real-world scenario data, not parameter count alone.
What is Goldman's valuation — and what is the logic?
Goldman maintained its 12-month target of HK$670, derived via a sum-of-the-parts methodology — valuing each of Tencent's business segments separately and adding them up.
At the time of publication, the share price was HK$455.20, implying roughly 47.2% upside.
In plain terms = Goldman believes the market is significantly underpricing Tencent's AI business. If the closed-loop strategy gains traction, that value should flow through to the overall valuation.
What are the risks to watch?
Goldman explicitly flagged that rising capex intensity and sustained AI investment will weigh on near-term profit growth.
Specifically, the bank forecasts year-on-year EPS growth of just 4% in Q3 2026 and 0% in Q4 2026.
This means → the key variables the market will track are Hunyuan's iteration pace, WorkBuddy user metrics, and the rollout of WeChat's Xiaowei assistant — whether they can sustain the mid-term valuation while earnings growth slows.
市场有风险，内容仅供研究参考，不构成投资建议。