Tencent launched the open-source Hy4 Preview generative AI model, presenting it as a productivity model aimed at programming, data analysis, document work, and scientific research. The model contains 770 billion parameters, while only 49 billion parameters are activated during each operation through a Mixture-of-Experts architecture, with a context window exceeding one million tokens.
The long context window enables the model to handle large amounts of code or documents within a single session, a capability directly related to tasks requiring an understanding of multiple files or the tracking of lengthy workflows. However, Tencent did not provide independent results or external comparisons proving the model’s performance in different operating environments.
Internal Results Surpass Two Competing Models
According to Tencent’s data, Hy4 Preview scored an average of 2.99 points out of 4 in a blind test conducted by the company, involving 163 Tencent experts and covering 203 engineering tasks. The model outperformed GLM-5.3 from Z.ai, which scored 2.92 points, and Kimi K3 from Moonshot AI, which scored 2.94 points.
These figures reflect an internal test designed and conducted by Tencent, and should therefore be read as an initial performance indicator rather than an independent market ranking. The material also does not clarify the details of the dataset, the distribution of tasks, or the full evaluation methodology, which are important considerations when comparing models with different sizes and architectures.
From Programming to Scientific Research
Tencent positions the model in the context of practical use, rather than only general conversational tasks. It says Hy4 Preview can understand code, develop plans for executing it, debug errors, and verify results, with particular attention to the quality of the interfaces it creates and the experience of interacting with them in frontend development tasks.
The model also enables data analysis and financial analysis, work on multiple documents, and the creation of documents, spreadsheets, and presentations. The company also mentions the ability to create playable game prototypes from a natural-language prompt, then develop them through multiple interactions with game engines.
The areas highlighted by Tencent include AI research, molecular dynamics simulation, condensed-matter physics, and fundamental mathematics. According to the available material, these capabilities remain company claims about the model’s uses and performance, rather than results published by independent research organizations.
Using the Model to Develop Itself
Tencent says Hy4 Preview helped improve some stages of its development, including training methods, data strategies, evaluation systems, and low-level operators. According to the company’s description, the model proposed different methods, conducted experiments, and then helped build subsequent iterations based on the resulting outcomes, code, logs, and notes.
Tencent describes this as an early loop of iterative self-improvement. The model was also used to analyze bottlenecks in the inference architecture and propose changes to operator fusion and communication optimization. The company says these efforts increased end-to-end productivity by 31.8% compared with the initial baseline.
Availability, Cost, and Current Limitations
Hy4 Preview can be used through Tencent products, including WorkBuddy, CodeBuddy, Yuanbao, and ima. Programmatic access is available through Tencent Cloud TokenHub and OpenRouter, and the company has also shared the model through Hugging Face with the open-source community.
- Price per million input tokens: $0.834.
- Price per million output tokens: $2.501.
- Price per million cached tokens: $0.042.
- WorkBuddy and CodeBuddy users receive free access for two weeks after launch.
Why does this announcement matter? Hy4 Preview combines a large total size, a mixture-of-experts architecture, and a context window exceeding one million tokens, along with open-source availability and publicly disclosed API pricing. This could make it an important option for teams working with long codebases or documents. At the same time, Tencent describes it as a preview release and says it will use user feedback to improve it and launch new models in the Hy4 family soon; therefore, the reliability of its results beyond internal tests, the stability of the initial release, and questions concerning licensing and real-world operation remain issues requiring verification before broad adoption in production environments.