CICC has released a research report stating that Tencent Holdings (00700) has officially launched and open-sourced its Hunyuan Hy4 preview language model, with a full version of Hy4 expected to follow. Additionally, the Workbuddy feature is now offering a two-week free trial. The firm maintains its "Outperform" rating and a target price of HK$666 on the stock.
The report highlights that the Hy4 preview employs a Mixture-of-Experts (MoE) architecture, featuring 770B total parameters, 49B activated parameters, and a 1M token context length. This marks a significant expansion over Hy3 in terms of model size, context length, and data scale. A key highlight is the breakthrough in scientific research capabilities, with several achievements demonstrating notable relative leadership in the AI4Science domain. Software engineering capabilities have also been enhanced, as the model utilizes MCP access to build game demos and interactive comics entirely from conversational prompts, validating its ability to understand, plan, debug, and verify long-horizon development tasks. On pricing, the input cost is RMB 6 per million tokens, output is RMB 18 per million tokens, and cache hits are priced at RMB 0.3 per million tokens, keeping it at a low level among comparable first-tier open-source models. The model is currently deployed in the Workbuddy series, Yuanbao, and ima products, and has also been listed on Tencent Cloud's TokenHub, accelerating the accumulation of high-quality data for the next phase of reinforcement learning. The official release notes acknowledge that long-context thinking in complex tasks and tendencies toward excessive self-verification remain known issues, and the market should watch for the subsequent full version launch.
The continuous leap in model capability is expected to improve investor confidence regarding return on investment (ROI). Since infrastructure was rebuilt in February of this year, Hunyuan has iterated on major versions on average every two months, significantly accelerating the development pace. The firm believes that as the bidirectional data loop between the model and agents gradually matures and the path to product commercialization becomes clearer, market confidence in the certainty of Tencent's AI investment returns is likely to strengthen. CICC remains optimistic about the company's strategic execution and delivery capabilities in the long-term AI赛道, maintaining its view that Tencent holds a strong probability of long-term success.
Comments