Tencent has open-sourced its 770-billion-parameter Hy4 preview model, giving developers access to a one-million-token-context AI system built for extended coding, research and knowledge-work tasks.
Tencent has released Hy4 preview as an open-source model, making the 770-billion-parameter system available through public repositories alongside Tencent Cloud, OpenRouter and Tencent products. The model has 49 billion active parameters and a one-million-token context window.
Tencent describes Hy4 preview as its most capable model so far, claiming its largest generation-over-generation improvement. The model targets long-horizon software engineering, document-heavy office work, game development and scientific research.
For software engineering, Hy4 preview can understand, plan, debug and verify work across extended sessions, including improving front-end visual output. It can also turn scattered files into documents, spreadsheets and presentations, while handling data analysis, equations and financial models. Tencent says it can create playable game prototypes from one prompt and refine complex projects across multiple turns.
The model is also positioned for more autonomous R&D workflows. Tencent says Hy4 preview managed several Codex sessions in parallel, changing its research direction as results arrived. In a small-model post-training task, it reportedly coordinated multiple evaluation targets and beat independently operating Codex across all eight benchmarks.
A blind internal evaluation by 163 Tencent experts across 203 engineering tasks gave Hy4 preview an average score of 2.99, versus 2.92 for GLM 5.3 and 2.94 for Kimi K3.
Tencent trained the model using work-related data from its software engineers, game developers, finance analysts and security experts. It acknowledges that the early version can over-reason and over-verify complex tasks.
