Tencent has stepped deeper into China’s rapidly changing artificial intelligence race with its new Hy4 Preview model. The company says Tencent has developed the model for practical work including coding, research, financial analysis, office tasks, and other demanding workloads.
The announcement matters because Tencent is not simply launching another experimental chatbot. Hy4 Preview is being presented as an open-source foundation model with a huge context window and a mixture-of-experts design. More importantly, Tencent says its model narrowly performed better than Z.ai’s GLM-5.3 and Moonshot AI’s Kimi K3 during internal testing.
That claim needs some context, though. The comparison came from Tencent’s own evaluation rather than a completely independent benchmark, so it would be too early to call Hy4 the undisputed leader. Still, the numbers show how quickly Chinese AI companies are catching up with each other.
Tencent Introduces Hy4 Preview
Tencent released the preview version of Hy4 on August 28, giving developers and researchers another open model to experiment with. The company says the system has been built around practical tasks rather than being focused only on conversational responses.
Hy4 Preview comes with 770 billion total parameters, while around 49 billion parameters are activated for an individual request. That mixture-of-experts approach allows the model to be very large without activating every parameter for every task.
The model also supports a context window of up to one million tokens. That is a particularly useful feature for developers and professional users working with large documents, lengthy codebases, research material, or complicated projects.
A large context window essentially allows an AI system to process much more information during one interaction. It does not automatically make a model better at every task, but it can make long-form work considerably more practical.
Internal Tests Put Hy4 Ahead
The most interesting part of the announcement is Tencent’s comparison with rival Chinese AI models. In a blind evaluation involving 163 experts and 203 engineering tasks, Hy4 Preview received an average score of 2.99 out of four.
Z.ai’s GLM-5.3 reportedly scored 2.92, while Moonshot AI’s Kimi K3 reached 2.94 in the same evaluation. That gives Hy4 a narrow lead over both models, rather than a dramatic performance advantage.
The difference is small enough that users should not interpret the result as proof that Hy4 is better at everything. Testing methodology, task selection, model versions, prompting methods, and evaluation conditions can all influence results.
There is another important point here. Tencent conducted the evaluation itself, meaning the numbers should be treated as company-reported results. Independent testing will be more useful for understanding how Hy4 compares across a broader range of real-world applications.
Coding Is A Major Focus
Coding appears to be one of the areas Tencent wants Hy4 Preview to handle seriously. The company says the model can help with understanding software projects, planning development work, debugging problems, and validating code.
That puts Hy4 directly into an increasingly competitive market where AI models are being used as coding assistants rather than simple question-answering tools.
Tencent has also demonstrated the model working on tasks such as creating game prototypes and building 3D websites from natural-language instructions. These examples point toward a broader direction for AI assistants, where users describe an outcome and the model handles multiple stages of execution.
This is important because modern AI competition is moving away from simple benchmark scores. Companies increasingly want models that can actually complete useful workflows with fewer human interventions.
Huge Context Window Adds Flexibility
The one-million-token context window is another major feature of Hy4 Preview. For ordinary chatbot conversations, most users will never need anything close to that amount of context.
Professional applications are different.
A developer could potentially provide extensive project information, multiple files, documentation, and related instructions within one session. Researchers could work with long technical documents, while financial teams could process large collections of information without repeatedly breaking everything into smaller pieces.
Of course, a large context window does not guarantee perfect understanding. AI models can still miss important details or produce incorrect conclusions when given huge amounts of information.
Still, having the ability to handle such a large amount of context gives Tencent more flexibility when positioning Hy4 as a productivity model.
Hy4 Was Used During Development
One of the more unusual claims from Tencent concerns the model’s own development process. The company says Hy4 participated in experiments related to training methods, data strategies, and evaluation systems.
According to Tencent, the results from those experiments were then used to improve later rounds of model optimization. The company has described this approach as an early form of a recursive self-improvement loop.
The idea is worth watching because AI companies are increasingly trying to use AI systems to improve the process of building future AI systems.
That does not mean Hy4 independently redesigned itself or became fully autonomous. Human researchers remained involved in the development process. The more accurate description is that Tencent used the model as part of an iterative research and optimization process.
Z.ai And Moonshot Remain Strong Rivals
Hy4’s arrival comes at an unusually competitive moment for China’s AI industry. Z.ai has recently attracted attention with its GLM models, while Moonshot AI has pushed its Kimi family into increasingly demanding reasoning and coding workloads.
Z.ai‘s GLM-5.3-Flash recently appeared anonymously under the name Ox Alpha before the company confirmed that it was behind the model. That episode generated considerable attention because developers were impressed by the model’s coding and agentic capabilities.
Moonshot AI, meanwhile, has been promoting Kimi K3 as a very large open-weight model. Reuters reported that the company is also discussing potential revenue-sharing arrangements with major cloud providers for hosting K3.
So Tencent is entering a market where established Chinese AI specialists are already moving extremely quickly.
Open Source Changes The Picture
Tencent is also making Hy4 Preview available as an open model. The company released its model weights and is making the system accessible through several Tencent products and services.
Those include CodeBuddy, WorkBuddy, Yuanbao, and ima, while developers can also access the model through APIs. Tencent has additionally offered free access to Hy4 on WorkBuddy and CodeBuddy for a limited period.
Open availability can be strategically important because developers can test the model directly rather than relying entirely on company demonstrations.
It also gives Tencent a chance to build an ecosystem around Hy4. If developers begin integrating the model into coding tools, business software, research workflows, and other applications, Tencent could gain much more than attention from a single model launch.
Independent Benchmarks Still Matter
Despite the strong internal results, Hy4 Preview still needs broader testing before its position becomes clear.
Reports indicate that results against other Chinese models can vary on third-party benchmarks. Hy4 has shown progress, but performance is not identical across every test or workload.
This is fairly normal for modern AI models. One system may perform extremely well in coding while another performs better in mathematics, reasoning, document analysis, or agentic tasks.
The strongest conclusion right now is that Tencent has produced a significantly improved model that appears competitive with leading Chinese systems.
Calling it the overall best model would require more independent evidence.
Tencent Wants A Bigger AI Role
For Tencent, Hy4 is about more than one model release. The company has enormous existing platforms across gaming, communication, advertising, cloud services, and productivity software.
That gives Tencent plenty of places where a stronger AI model could eventually be integrated.
The company’s recent AI spending also shows that it is taking the technology race seriously. Reuters reported that Tencent increased capital expenditure on AI projects and computing substantially during the quarter ended June.
The company can potentially connect its AI research with products already used by millions of people. That could become an important advantage if Hy4 continues improving.
What Hy4 Means For AI Users
For developers, Hy4 gives another powerful open model to test and potentially integrate. Its large context window could be especially useful for software development, research, document processing, and complicated enterprise workflows.
For businesses, the bigger question will be reliability and operating cost. A model can score highly in testing and still struggle when deployed across thousands of real users.
For the wider AI market, however, the release sends a clear message. Chinese companies are not slowing down their model development efforts.
Tencent’s latest move also shows that competition is becoming increasingly crowded. Z.ai, Moonshot AI, Alibaba, Tencent, and other companies are all trying to produce models capable of handling sophisticated professional tasks.
The Bigger AI Competition Continues
Tencent Hy4 Preview is another sign that China’s AI industry is moving rapidly toward larger, more capable, and increasingly open models. Its reported 2.99 score in Tencent’s internal engineering evaluation gives the company a useful headline, especially after beating GLM-5.3 and Kimi K3 by relatively small margins.
Still, the most important test will happen outside Tencent’s own evaluation environment. Developers, researchers, independent benchmark creators, and enterprise users will determine whether Hy4 can consistently deliver strong results in practical situations.
The model’s 770-billion-parameter architecture, one-million-token context window, open availability, and focus on coding and productivity make it a serious addition to the current AI landscape. Tencent now has an opportunity to turn that technical progress into widespread adoption across its ecosystem.
Read More :- RedTeaDetoxCleanse.com