Tencent's next-generation large language model Hy4 previewWe released it and deployed the model weights. The total number of parameters is 770 billionand the context window is Exceeding 1 million tokens Although it's a large model, what stands out most about this announcement is its intended use rather than its size.
Tencent introduced the Hy4 preview as a model tailored for productivity tasks—such as coding, document and data analysis, game development, and scientific research—rather than general conversation. WorkBuddy·CodeBuddy·Yuanbao·ima We integrated it with our own services and also made the API available to external developers.
However, it’s an early model with “preview” in its name, and most of the performance comparisons are based on evaluations conducted by Tencent. The fact that it can process 1 million tokens and the fact that it always accurately understands long documents are two different matters, so we need to look at the actual usage results together.
| It is not a model that uses all 770 billion at once.
The Hy4 preview is MoE (Mixture of Experts) Use the architecture. It's a method where you place multiple specialized modules within the model and select only the ones needed to process the input tokens.
Although there are 770 billion parameters in total, the number of parameters activated when processing a single token is approximately 49 billionHey. You could say this design aims to secure the massive knowledge capacity of a large model while reducing the burden of calculating every parameter each time.
That doesn't mean, however, that it can be driven as easily as a smaller model. Since the total weight set that has been released is very large, running it directly still requires a significant amount of GPU memory and storage space, as well as a distributed inference environment.
| What is the context for over 1 million tokens used for?
The context window refers to the range of input that a model can reference in a single task. Supporting over 1 million tokens increases the likelihood of handling large code repositories, multiple volumes of documentation, and lengthy research papers and work records within a single task.
It can be particularly useful when comparing the contents of different documents or reviewing old code alongside new changes. Tencent also cited understanding long blocks of code, creating documents, spreadsheets, and presentation materials, and cross-document collaboration as key use cases.
However, the maximum input length does not guarantee a certain score. Separate verification is needed to ensure that no specific details in the middle of the document are overlooked, to determine how much response latency and costs increase as the input lengthens, and to verify that sources are properly cited.
| Focus on Productivity, from Coding to Financial and Scientific Research
Tencent explained that it trained the Hy4 preview for use in software engineering, office work and data analysis, game development, and scientific research. In the area of coding, the company stated that it enhanced the model’s ability to understand the structure of long-term projects, develop plans, debug, and validate code.
In administrative work, focus on the process of analyzing various documents and then compiling the results into documents, tables, and presentation materials. Although financial analysis is also a major area of application, the analysis generated by the model is prone to errors, so human verification is required for actual investment decisions or important decisions.
In the field of science, examples included AI research, molecular dynamics, condensed matter physics, and problems in pure mathematics. These are the development goals set by Tencent and the results of its own evaluations; they do not imply that verification has been completed across all specialized fields.
| Public weights and our own services and APIs simultaneously
The model weights for Hy4 Preview have been released via Hugging Face. Developers can download them directly to test them, and they are available in Tencent products such as WorkBuddy, CodeBuddy, Yuanbao, and ima.
For users who find it difficult to manage on their own Tencent Cloud TokenHubWow OpenRouterIt also provides an API. The price announced by Tencent is per 1 million tokens $0.834, Output $2.501, Enter cache hit $0.042Hey.
Offering both public weights and a paid API gives businesses more options. Organizations that want to process sensitive data within their own environments can consider operating the system themselves, while developers looking to reduce their infrastructure burden can opt for the API.
| Self-Assessment
Tencent has in-house experts 163 peopleThis actual engineering work 203In a blind test evaluating [...], Hy4 Preview scored an average of [...]. out of 4 2.99He revealed that he had received it. They say the GLM 5.3, which was included in the comparison, scored 2.92, while the Kimi K3 scored 2.94.
The company explained that the model was also involved in the development process of analyzing bottlenecks in the inference system and optimizing operator combinations and communication methods. Through this process, end-to-end throughput was 31.8% They also presented figures showing that it had increased.
However, we need to consider separately the fact that this was an internal test designed by Tencent, including its evaluation criteria and environment. Competitiveness will become clearer once independent organizations conduct replication studies and results from various languages and work environments are accumulated.
| The Limitations Implied by the Name “Preview”
Tencent noted that, when faced with complex questions, the model may take longer than necessary to generate a response or over-validate its own answers. This means that long context and strong reasoning capabilities can actually result in slow responses or unnecessary token usage.
The significance of the Hy4 preview lies in the fact that a major Chinese tech company has integrated a massive open model into both its products and APIs simultaneously. Going forward, it will be necessary to verify the actual operating costs of the published weights, their performance across multiple languages—including Korean—their accuracy with long documents, and the results of independent evaluations.
References
Tags
#Tencent #Tencent #Hy4preview #Hunyuan #Open-Source AI #Large Language Models #MoE #1M Tokens #AI Model #CodingAI #Generative AI #AI Model Services
