Tencent has released its next-generation large language model Hy4 Preview as open source. Focused on coding, office applications and scientific research, the model has 770 billion parameters; with the Mixture-of-Experts approach, 49 billion parameters are activated per operation. Its context window exceeds 1 million tokens.
In a blind test conducted by Tencent with 163 experts and 203 engineering tasks, Hy4 Preview scored 2.99 out of 4; Kimi K3 scored 2.94, while GLM-5.3 remained at 2.92. The company stated that the results were based on its own internal evaluation.
The model can be used for long-context software development, data and financial analysis, document, spreadsheet and presentation generation, game prototyping, and solving scientific problems. Tencent also announced that Hy4 Preview contributed to training, data, evaluation and operator optimization, increasing end-to-end capacity by 31.8% by eliminating bottlenecks in its own infrastructure.
The model is available through WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, OpenRouter and Hugging Face. API prices were set at 0.834 dollars per 1 million input tokens, 2.501 dollars per 1 million output tokens, and 0.042 dollars per 1 million cached tokens.
Why it matters
The open-source release of Hy4 Preview transforms the model from a service limited to Tencent’s own products into an option that developers and researchers can examine across different use cases. Its context capacity of more than one million tokens enables the model to process a broader range of content at once, particularly for teams working with long software projects, extensive documents and large datasets. Access through WorkBuddy, CodeBuddy, Tencent Cloud TokenHub, OpenRouter and Hugging Face, along with the disclosure of API prices, allows usage decisions to be evaluated based not only on technical features but also on cost. However, since the claim of performance superiority is based on Tencent’s own internal blind test, it remains unclear whether the same result would be obtained in different and independent evaluations.