Bernstein Comments on Tencent's Hy3 Model: Steady Improvement in AI Capabilities, Clearer Path to Agent Commercialization

The official version of Tencent's large AI model, Hy3, was released on July 6, just about 10 weeks after the preview version went live. Bernstein subsequently issued a report, maintaining its "Outperform" rating for TENCENT. The firm believes that while Hy3 is not a frontier model, its performance is sufficient to support the deployment of Agent applications within the WeChat ecosystem, and that market concerns regarding token costs and capital expenditures may be overstated.

Hy3 retains the architecture of 295 billion total parameters and 21 billion activated parameters, but features significant optimizations in key capabilities: improved stability in tool invocation, a reduction in hallucination rates from 12.5% to 5.4%, enhanced memory for multi-turn conversations, and overall reasoning capabilities reaching levels comparable to GLM-5.1. However, there are still trade-offs in heavy-duty programming and general reasoning. The GQA architecture can be further iterated in the future through KV Cache compression technologies such as MLA and DSA.

Bernstein pointed out that Hy3 is the first complete achievement of Tencent's new AI team, and its performance improvements verify that the model pre-training, reinforcement learning, and evaluation systems are on the right track. Public benchmark tests show that Hy3's overall performance is significantly ahead of MiniMax M3. The next-generation model is expected to launch as early as late 2026 or early 2027, potentially introducing a completely new pre-training architecture.

Analysts believe that Tencent's AI strategy is shifting from model R&D to commercial implementation. The core metric for measuring return on investment in the future will be the scale of Agent transactions, rather than chatbot usage. The path for Agent services built around the WeChat ecosystem is becoming clearer, which is key to the long-term value of Tencent's AI.

Agent Commercialization Is More Important Than Model Competition

Rather than the model capabilities themselves, Bernstein is more focused on the commercialization path of Tencent's Agent business.

The report noted that while US internet platforms recently saw a surge in Agent-based e-commerce that quickly cooled down, Chinese internet platforms naturally possess closed-loop transaction systems. A large volume of user transactions already exists within super-apps like WeChat, making the foundation for Agent commercialization more mature.

Bernstein expects that Tencent will not rush to charge ordinary users, but is more likely to adopt a B-side charging model. Specifically, merchants subscribing to Agent services will gain access to more AI-driven traffic, richer AI tools, and deeper integration within the WeChat ecosystem; non-subscribing merchants will receive correspondingly less traffic and resources. This means Tencent's future AI revenue is more likely to come from merchant marketing and service fees rather than direct consumer subscriptions.

Concerns Over Token Costs May Be Exaggerated by the Market

As global internet companies continue to increase investment in AI infrastructure, Tencent's capital expenditures have also become a focus for investors.

Tencent has previously guided that capital expenditures will increase quarter by quarter in 2026, leading to corresponding increases in depreciation and amortization expenses. However, Bernstein believes that the market's current concerns about AI token consumption costs are somewhat overinterpreted. The report points out that a typical chatbot conversation usually consumes only hundreds of tokens, whereas genuine Agent transactions often require tens of thousands of tokens.

This means that Tencent's token consumption will see an exponential increase only when the volume and Gross Merchandise Value (GMV) of Agent transactions grow on a large scale. From a business model perspective, there may be a time lag between user activity growth and revenue growth, but historical experience shows that commercialization is ultimately a matter of time, not feasibility.

WeChat Agents Still Face Internal Coordination Challenges

However, Bernstein also pointed out that there remains an important unresolved issue in Tencent's AI strategy.

The report stated that while Tencent's reorganized AI team has made significant progress in model training infrastructure, it is understood that this team has not yet obtained access to WeChat data. Meanwhile, the upcoming WeChat AI Agent is being developed primarily by the WeChat team independently.

Analysts believe that this fragmentation caused by the organizational structure will likely be resolved through management coordination eventually. However, why deeper synergy has not yet been achieved between the two sides will remain an important observation point for investors monitoring the execution of Tencent's AI strategy in the future.