CICC: Tencent Holdings (00700) Open Source Releases Hy4 Preview Model, Jumping Capabilities Are Expected to Improve Investment ROI Concerns

The Zhitong Finance App learned that CICC released a research report stating that Tencent Holdings (00700) officially released and open source the hybrid Hy4 preview language model, which will later replace the official version of Hy4; at the same time, Workbuddy released a 2-week free trial. Maintain the “Outperform the Industry” rating and target price of HK$666.

According to the report, Hy4 preview uses a mixed expert architecture. The total number of parameters is 770B, activation parameters are 49B, and the context length is 1M. Compared with Hy3, it is significantly expanded in terms of model size, context length, and data scale. The key highlights are breakthroughs in scientific research capabilities. Many of its achievements have significant relative leadership on the AI4Science circuit; software engineering capabilities have been enhanced to interactively build game demos and interactive comics from scratch through MCP access to verify understanding, planning, debugging and verification capabilities for long-term development tasks; on the price side, 6 yuan/million tokens are input, 18 yuan/million tokens are output, and the cache hits 0.3 yuan/million tokens, maintaining a low position in the first-line open source model with the same parameters.

Currently, the model has been launched in products such as the Workbuddy series, YuanBao, and ima, and has been launched on Tencent Cloud TokenHub to accelerate the accumulation of high-quality data for the next phase of reinforcement learning (officially believes that long-term thinking in complex tasks and excessive self-verification tendencies are still known problems). It is recommended to pay attention to subsequent official releases. A continued jump in model capability is expected to improve investment ROI concerns. Since the infrastructure was rebuilt in February of this year, mixed yuan has been stacked on average once every two months, and the pace has accelerated markedly.

The bank believes that as the two-way data between the model and agent gradually runs through a closed loop and the path to commercialization is becoming more clear, the market's previous certainty about the return on Tencent's AI investment is expected to continue to increase. It is optimistic about the company's strategic execution and ability to deliver on the AI long-term circuit, and still maintain a strong long-term victory rate.