Tencent’s newest AI model did more than emerge from the company’s development process. Tencent says the model actually helped with parts of that process.
The company released Hy4 preview on Aug. 28, calling it its most capable model yet for software engineering, office work and scientific research. Hy4 has 770 billion total parameters, 49 billion active parameters and a context window exceeding 1 million tokens.
Tencent has open-sourced Hy4 and made it available through Tencent Cloud and OpenRouter, while integrating it into products such as CodeBuddy, WorkBuddy, Yuanbao, and ima. But one of the more significant details is how Tencent says it built the model: Hy4 was used to propose experiments, analyze results and identify ways to make its own inference infrastructure more efficient.
The focus is practical work. In coding, Hy4 is designed to handle long-running development tasks, including planning, debugging and validation. It can also generate playable game prototypes from natural-language prompts and work with game engines.
For office and analytical work, Tencent says the model can process information spread across multiple files and turn it into documents, spreadsheets and presentations. It has also been trained for financial analysis and data-heavy workflows.
A narrow lead over rivals
Tencent’s internal blind test gives Hy4 a slight edge over two major Chinese competitors.
In an evaluation involving 163 Tencent experts and 203 engineering tasks, Hy4 scored 2.99 out of 4. GLM-5.3 from Z.ai scored 2.92, while Moonshot AI’s Kimi K3 scored 2.94. Hy4 recorded 46.8% wins against GLM-5.3 and 51.2% against Kimi K3.
The results should be viewed carefully, however, because the comparison was conducted internally by Tencent rather than by an independent testing organization.
Why it matters: Hy4 helped improve itself
Tencent says Hy4 participated in its own development by proposing approaches, running experiments, and using the resulting code, logs, and feedback to guide subsequent experiments.
The model also analyzed bottlenecks in its inference system and helped optimize areas such as operator fusion and communication. Tencent says those changes increased end-to-end throughput by 31.8% compared with its baseline.
That does not mean Hy4 can independently redesign itself without human oversight. But it points to a potentially important shift: models becoming tools to improve the expensive processes used to train and operate future models.
Cheap access, with some catches
Tencent is pricing Hy4 at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens. That pricing could make its large context window attractive to developers working with lengthy codebases, documents, or research materials.
But Hy4 is explicitly a preview. Tencent acknowledges that it can spend too long reasoning through difficult tasks and sometimes over-verify its own work. The model also lacks multimodal vision capabilities.
Hy4 also shows how quickly Tencent is compressing its AI release cycle. After rebuilding its AI infrastructure earlier this year, the company moved from Hy3 preview to the full Hy3 model and now Hy4 preview, with another iteration already planned.
For developers and AI solution providers, the bigger trend may be what happens between those releases. If models increasingly help researchers test code, diagnose infrastructure bottlenecks, and optimize inference, AI companies could shorten the development cycle for the next generation of models.
Also read: Stripe’s $7.5 billion OpenRouter deal could reshape how businesses access, route, and manage spending across AI models.