Tencent's latest artificial intelligence model participated in its own creation, a development that the company disclosed when it released Hy4 preview on Aug. 28. The Chinese tech giant says the model suggested experiments, examined outcomes, and pinpointed opportunities to make its own processing infrastructure work faster. Tencent calls Hy4 its strongest model yet for software engineering, workplace tasks, and scientific research, and it has made the system available as open-source software through Tencent Cloud and OpenRouter.
Hy4 contains 770 billion total parameters with 49 billion active parameters and can handle context windows larger than 1 million tokens. The model costs $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens. In an internal evaluation run by Tencent with 163 company experts and 203 engineering assignments, Hy4 earned a score of 2.99 out of 4. That edged out GLM-5.3 from Z.ai, which scored 2.92, and Moonshot AI's Kimi K3, which scored 2.94. Hy4 won 46.8% of comparisons against GLM-5.3 and 51.2% against Kimi K3. The company has built Hy4 into several products including CodeBuddy, WorkBuddy, Yuanbao, and ima, and designed it to manage extended coding projects involving planning, debugging, and validation, as well as create playable game prototypes from text instructions.
According to Tencent, Hy4 proposed approaches, ran experiments, and used the resulting code, logs, and feedback to steer follow-up experiments during its own development. The model examined congestion points in its processing system and helped refine areas such as operator fusion and communication. The company acknowledges that Hy4 is explicitly a preview version, can take too much time reasoning through challenging assignments, and occasionally over-checks its own output. The model also lacks capabilities for processing images or other visual information. The comparison with rival Chinese models was carried out by Tencent itself rather than by an outside testing body.
Tencent says those optimizations increased end-to-end throughput by 31.8% compared with its starting point. That doesn't mean Hy4 can redesign itself without people watching over the process, but it signals a potentially meaningful change: models becoming instruments to enhance the costly procedures used to train and run future models. For office and analytical work, Tencent says the model can digest information scattered across many files and convert it into documents, spreadsheets, and presentations, and it has received training for financial analysis and workflows heavy on data. The low pricing could make Hy4's large context window appealing to developers handling lengthy codebases, documents, or research files. Hy4 also demonstrates how rapidly Tencent is compressing its AI release schedule—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 version already on the calendar.
The bigger pattern may be what unfolds between those releases. If models increasingly help researchers test code, diagnose infrastructure bottlenecks, and optimize processing, AI companies could shorten the development cycle for the next generation of models. That cycle compression could change the pace at which businesses need to evaluate and adopt new systems, particularly for organizations that depend on stable, predictable technology roadmaps. The line between model training and model deployment may also blur as self-improvement becomes a continuous rather than episodic process.