Following the previous financial report meeting announcement, the internal test interface of Tencent's new generation large model Hy4 has been officially leaked recently, which means its official release is imminent. From the leaked model selection interface, it can be seen that Tencent continues its established strategy of "self-developed models and integrated excellent open-source solutions in parallel," offering three core options: Hy4, positioned as an expert-level large model; Hy3, marked as a newly upgraded model suitable for most scenarios; and DeepSeek, focusing on reasoning and understanding.
As the focus of this upgrade, Hy4's overall positioning is clearly higher than its predecessor Hy3, and it is given expert-level capabilities to solve more complex business and technical challenges. In terms of technical evolution, the industry generally speculates that its significant performance improvement may come from two directions: one is directly scaling up the parameter size to the 1 trillion level to match industry flagship standards; the other is further deepening the reinforcement learning (RL) training based on the existing Hy3 architecture, thus significantly enhancing the core intelligence ceiling without blindly expanding the parameters.
Looking back at the Hy3 large model released this year, its total parameter count is 295B (activated parameters 21B, MTP layer parameters 3.8B), and it has performed well in high-demand tasks such as software development, office production, financial modeling, front-end design, and game creation. In a large-scale internal expert real work blind test, Hy3's overall performance even exceeded that of similar models. With the upcoming release of Hy4, which has a higher position and stronger performance, it is expected to fully take over as the core underlying support for the intelligent agent application WorkBuddy, further unleashing Tencent's application potential in AI intelligent agent implementation and productivity scenarios.
