Tencent open-sources Hy4 preview with 770B parameters and a 1M-token context
Tencent has open-sourced its Hy4 preview large language model, featuring 770 billion total parameters and a 1 million token context window, available through various platforms and APIs.
Intelligence analysis by Gemini 2.5 Flash
Tencent's new Hy4 preview LLM, with its massive parameter count and context window, is now open-source and accessible via WorkBuddy, CodeBuddy, Yuanbao, ima, and Tencent Cloud TokenHub/OpenRouter APIs. It targets productivity tasks like coding and data analysis, showing competitive performance in internal evaluations against models like GLM 5.3 and Kimi K3.
Imagine a super-smart computer brain called Hy4 that Tencent just shared with everyone. It's like a giant library with a million books it can read all at once to help you write computer code, do your office homework, or even make games. For a short time, you can try it for free, and it's really good at helping with tricky tasks, like a super-fast helper for your computer.
Analysis
Hy4 Preview
Tencent has officially launched and open-sourced its Hy4 preview, a next-generation large language model, marking a significant step in the company's AI strategy. This model boasts an impressive 770 billion total parameters, with 49 billion activated parameters, indicating a sophisticated architecture designed for efficiency and performance. A key feature is its context window, which exceeds 1 million tokens, allowing the model to process and understand exceptionally long inputs and conversations.
The Hy4 preview is specifically engineered to tackle a wide array of productivity tasks. These include complex coding challenges, general office work, in-depth data analysis, game development, and even scientific research applications. Its broad utility suggests Tencent aims for a versatile tool capable of assisting professionals across multiple demanding fields, potentially streamlining workflows and accelerating innovation.
770 Billion Parameters
The sheer scale of Hy4 preview, with its 770 billion total parameters, places it among the largest language models globally. While the article notes 49 billion activated parameters, this distinction is crucial; it suggests a sparse or mixture-of-experts architecture, where not all parameters are engaged for every computation, optimizing for both performance and computational cost. This design choice allows for a massive knowledge base while maintaining practical inference speeds.
Such a large parameter count typically correlates with enhanced capabilities in understanding nuance, generating coherent and contextually relevant text, and performing complex reasoning tasks. The model's ability to handle a 1 million token context window further amplifies these capabilities, enabling it to maintain long-term coherence and draw insights from extensive documents or codebases. This is particularly beneficial for tasks requiring deep contextual understanding, such as debugging large code projects or summarizing lengthy research papers.
WorkBuddy and CodeBuddy
Access to the Hy4 preview is being facilitated through several platforms, including the Chinese and international versions of WorkBuddy and CodeBuddy, as well as Yuanbao and ima. This multi-platform availability ensures a broad reach for developers and enterprises looking to integrate the model into their operations. For a limited two-week period, users of WorkBuddy and CodeBuddy will enjoy free access, encouraging widespread adoption and testing.
Beyond the free trial, API access is provided via Tencent Cloud TokenHub and OpenRouter, with a pricing structure of $0.834 per million input tokens and $2.501 per million output tokens. This tiered pricing model is standard for large language model APIs, balancing accessibility with the computational resources required. Internally, Hy4 preview demonstrated strong performance, achieving an average score of 2.99 out of 4 in a blind evaluation involving 163 experts and 203 engineering tasks, surpassing competitors like GLM 5.3 (2.92) and Kimi K3 (2.94). Tencent also reported that the model optimized its own training and inference systems, boosting end-to-end throughput by 31.8%.
Key points
- Tencent open-sourced its Hy4 preview large language model on August 28.
- The model features 770 billion total parameters, 49 billion activated parameters, and a 1 million token context window.
- It is accessible via WorkBuddy, CodeBuddy, Yuanbao, ima, and through Tencent Cloud TokenHub and OpenRouter APIs.
- Hy4 preview is designed for productivity tasks including coding, office work, data analysis, game development, and scientific research.
- Internal evaluations showed Hy4 preview outperforming GLM 5.3 and Kimi K3 in engineering tasks, and it improved Tencent's own system throughput by 31.8%.
The open-sourcing of Hy4 preview could significantly accelerate innovation in AI development, providing researchers and developers with a powerful tool to build more sophisticated applications. Its large context window and competitive performance suggest it could significantly boost productivity in coding, office work, and scientific research, fostering new advancements.
While powerful, the model's high parameter count might pose significant computational demands, potentially limiting its accessibility or increasing operational costs for broader adoption beyond the initial free period. The reliance on internal blind evaluations also means its real-world performance against a wider range of benchmarks remains to be fully validated by external parties.
