Bernstein expects Tencent Hy3 progress to continue, and maintains Outperform with an HK$780 target price

Overview

This report focuses on the investment implications following Tencent’s formal Hy3 release. Bernstein believes the Hy3 final model has achieved substantial improvements in key capabilities, validating Tencent’s AI team strategy in pre-training, reinforcement learning, and evaluation stack development. The report also discusses capital expenditures, token costs, agent commercialization paths, and data and team coordination issues between Weixin and the AI team.

Core views

The core view is that Hy3 is not a top-tier frontier model, but is sufficient for Tencent to advance agent services within its ecosystem; market concerns over AI token spend and capex are overly pessimistic; because China’s internet platforms inherently have transaction intent and a merchant base, Tencent’s AI agent monetization is more likely to materialize through merchant subscriptions and traffic distribution mechanisms. Bernstein maintains a Tencent Outperform rating and an HK$780 price target.

Analysis framework

The report combines model-capability benchmark comparisons, ecosystem monetization-path analysis, cost-structure discussion, and platform-organization synergy observations. The authors compare Hy3 with GLM-5.1 and Minimax M3 capabilities and, considering differences in AI monetization environments between China and the U.S., assess potential return paths for Tencent’s AI investment.

Methodology notes

• AI model assessmentBenchmark comparison

  Measure Hy3’s progress through indicators such as tool-call stability, hallucination control, multi-turn retention, and reasoning capability.

  The report says Hy3 improved hallucination control from 12.5% to 5.4%, with the multi-turn retention MRCR comparison at 42.9% vs 75.1%, and overall reaching partial GLM-5.1-level capability.

• Cost and monetization analysisToken spend economics

  Distinguish low-token-consumption chatbots from high-token-consumption agent transactions and assess the conditions under which cost pressure emerges.

  The report argues that chat typically consumes hundreds of tokens, while agent transactions may require tens of thousands of tokens; Tencent’s token consumption would increase materially only when agent transaction volume and GMV rise sharply.

• Platform monetization frameworkMerchant-side monetization

  Position merchants, rather than consumers, as the primary payers for AI agent services.

  The report argues that Chinese internet platforms already have abundant transaction intent, and Tencent can monetize by charging merchants for AI traffic, tools, and WeChat ecosystem access, whether open or controlled.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• 00700.HK / Tencent Holdings Ltd

  The report covers the company and the listed security

  Strengths

  The Hy3 final release has improved key capabilities, and Tencent has the WeChat ecosystem, merchant connectivity, and pre-existing transaction intent within China internet platforms.

  Weaknesses

  Hy3 is still not a frontier model, and heavy coding as well as general reasoning are not its priority direction; there is further optimization room in GQA architecture.

  Comparison

  The report believes Hy3 reaches a GLM-5.1 level for Tencent’s required capabilities and is clearly ahead of Minimax’s M3 model.

  Risks

  Capex ramp, depreciation and amortization pressure, token spend concerns, insufficient coordination between Weixin data and the AI team, and uncertainty over the pace of monetization from agent transaction volume.

Key data

• RatingOutperformBernstein's investment rating for Tencent.
• Target priceHK$780.00Price target listed in the report.
• Close priceHK$452.00As of July 6, 2026, 700.HK close price.
• Implied upside73%Table shows Upside/(Downside).
• Market capitalizationHK$4,109.69bnMarket Cap shown in the report table.
• Enterprise valueHK$3,587.55bnEV shown in the report table.
• F26E revenues828,757mFinancial table shows Revenues F26E.
• F27E net income316,569mFinancial table shows Net Earnings F27E.
• F26E adjusted P/E13.1xValuation table shows Adjusted P/E.
• F27E EV/EBIT9.2xValuation table shows EV/EBIT.

Impact & implications

If Bernstein’s view is correct, Tencent's AI investment should be seen as ecosystem-level agent capability building rather than just a short-term cost burden. Hy3’s progress increases the credibility of Tencent’s future rollout of WeChat agents and merchant-side AI services; at the same time, capital expenditures and depreciation/amortization will still weigh on near-term earnings, but token cost pressure should be assessed together with the scale of agent transactions and commercialization progress.

Risks

• Tencent has guided that 2026 capex will rise quarter by quarter, which may create depreciation and amortization cost pressure.
• If agent transaction volume and GMV expand quickly, token consumption could increase significantly.
• Hy3 is still not a frontier model, and there is room for improvement in heavy coding, general reasoning, and KV cache compression.
• The report notes that the AI team does not seem to have full access to Weixin data, and the upcoming Weixin AI agent may be built independently by the Weixin team, making cross-team coordination a management issue that needs to be addressed.

What to watch

• Whether the next-generation Hy3 model will be launched by late 2026 or early 2027, and whether it includes a new pre-training iteration.
• The launch timing, feature scope, user interaction, and merchant onboarding approach of the Weixin AI agent.
• Actual changes in Tencent’s AI capex, depreciation/amortization, and token costs.
• Whether merchants are willing to pay for AI traffic, tools, and WeChat ecosystem connectivity.
• Whether management resolves the collaboration issue between the AI model team and Weixin data and product teams.