Powerful New Launch Challenges Moonshot and Z.AI in China’s Fast-Moving AI Race
Tencent Hy4 AI Model Enters the AI Race With a Major Upgrade
China’s artificial intelligence market has another major contender. Tencent has officially released and open-sourced its new Hy4 preview model, adding fresh pressure to rivals such as Moonshot AI and Z.AI. The launch is important because Tencent is not presenting Hy4 as a model built only for chatting. Instead, the company says it has been designed around practical work, including coding, office tasks, financial analysis, gaming and scientific research. (Tencent)
The Tencent Hy4 AI Model comes with 770 billion total parameters, while 49 billion parameters are activated for each token. It also offers a context window of more than one million tokens. In simple terms, this means the model can work with very large amounts of information in a single interaction. That feature could be useful for developers, researchers and businesses handling long documents or large projects. (Tencent)
Tencent’s launch also shows how quickly competition is growing among Chinese AI companies. Moonshot AI and Z.AI have already attracted attention with their Kimi and GLM model families. Now, Tencent is trying to gain ground with a model that combines a large architecture with a strong focus on real-world productivity.
What Makes Tencent Hy4 AI Model Different?
The Tencent Hy4 AI Model uses a Mixture-of-Experts, or MoE, architecture. This approach does not require every part of the model to work on every token. Instead, selected expert sections are activated for a particular task. Tencent says Hy4 has 770 billion total parameters but activates 49 billion parameters per token. (GitHub)
The model has 78 layers, with 256 routed experts and one shared expert in its MoE layers. For each token, eight routed experts plus the shared expert are activated. Tencent has also included a native multi-token prediction layer to support speculative decoding. (GitHub)
This design matters because bigger models can be expensive to run. An MoE structure can allow a company to build a very large model while limiting the amount of computation used for each individual token. Therefore, the real-world cost and speed of Hy4 will be just as important as its headline parameter count.
The Tencent Hy4 AI Model also uses Gated DeepSeek Sparse Attention and IndexCache, according to the company’s technical documentation. Tencent says these technologies help the model manage information more efficiently across its long context. (GitHub)
Key Tencent Hy4 AI Model Data
| Feature | Tencent Hy4 Preview |
|---|---|
| Total parameters | 770 billion |
| Activated parameters | 49 billion |
| Architecture | Mixture-of-Experts |
| Context window | More than 1 million tokens |
| Layers | 78 |
| Routed experts | 256 |
| Activated routed experts | 8 per token |
| Main focus | Coding, office work, research, gaming |
| Release | Open source |
| Tencent products | WorkBuddy, CodeBuddy, Yuanbao and others |
| API access | Tencent Cloud TokenHub and OpenRouter |
The figures above come from Tencent’s official model documentation and launch announcement. (Tencent)
Tencent Says Hy4 Edges Past Z.AI and Moonshot
One of the most interesting parts of the launch is Tencent’s comparison with competitors. The Tencent Hy4 AI Model was tested internally against Z.AI’s GLM-5.3 and Moonshot’s Kimi K3.
Tencent says its blind evaluation involved 163 internal experts and 203 engineering tasks. Hy4 received an average score of 2.99 out of 4.00. GLM-5.3 received 2.92, while Kimi K3 scored 2.94. (Tencent)
The difference is relatively small, which is worth highlighting. Hy4 did not win by a huge margin in Tencent’s internal evaluation. Instead, the results suggest that the three systems are operating in a closely contested area for the tested engineering tasks.
Tencent’s published results also show that Hy4 recorded a 46.8% win rate, 12.8% tie rate and 40.4% loss rate against GLM-5.3. Against Kimi K3, the model recorded a 51.2% win rate, 7.9% tie rate and 40.9% loss rate. (GitHub)
These numbers should be viewed carefully because the evaluation was conducted by Tencent itself. They are useful for understanding the company’s position, but they should not be treated as a final independent ranking of all AI models.
That distinction is important for readers. The Tencent Hy4 AI Model appears competitive based on the available evidence, but broader independent testing will be needed to determine how it performs across a wider range of tasks.
A One-Million-Token Context Window Could Be a Big Advantage
Another major feature of the Tencent Hy4 AI Model is its context window of more than one million tokens. A context window determines how much information an AI model can process and keep available during an interaction.
For normal questions, users may never need such a large window. However, the feature can become valuable when working with large software projects, long reports, research papers, company records or multiple documents.
For example, a developer could potentially give an AI assistant a large project and ask it to understand the relationships between different files. Similarly, a business user could work with lengthy reports and spreadsheets without constantly splitting information into smaller sections.
However, a large context window does not automatically mean better answers. The model must still find the right information, understand it correctly and avoid making mistakes. Therefore, actual performance on long-context tasks will be more important than the headline one-million-token figure.
Tencent Targets Real Business Work
The Tencent Hy4 AI Model has been built with practical work in mind. Tencent says it worked with internal experts in software engineering, gaming, finance and security while creating training data and testing workflows. (Tencent)
In software development, Tencent says Hy4 is stronger at understanding projects, planning tasks, debugging code and checking results. The company also highlights improvements in front-end development, including visual quality and user interaction.
The model is also designed for office and analysis work. Tencent says Hy4 can turn information spread across different files into documents, spreadsheets and presentations. It can also handle data analysis, equations and financial models.
Gaming is another area where the Tencent Hy4 AI Model could have an advantage because Tencent already has deep experience in the gaming industry. According to its technical documentation, Hy4 can help create playable game prototypes from prompts and continue refining projects through multiple interactions. (GitHub)
Scientific research is another target. Tencent says the model has improved reasoning and problem-solving capabilities for areas such as AI research, molecular dynamics, condensed matter physics and mathematics.
Open-Source Release Changes the Picture
The Tencent Hy4 AI Model is not being offered only as a closed product. Tencent has released the model as open source and made it available through several channels.
Users can access Hy4 through Tencent products including WorkBuddy and CodeBuddy, while API access is available through Tencent Cloud TokenHub and OpenRouter. Tencent also said Hy4 would receive a limited free period on WorkBuddy and CodeBuddy after launch. (Tencent)
This open approach could help Tencent reach developers outside its own product ecosystem. Developers can study the model, experiment with it and potentially adapt it for different applications, depending on the model license and deployment requirements.
The Tencent Hy4 AI Model is also available through technical repositories and model platforms, giving researchers and developers more direct access to its architecture and deployment information. (GitHub)
For the wider AI market, open releases can increase competition because developers have more choices. That can encourage companies to improve performance, reduce costs and build more useful AI tools.
Tencent Is Spending Heavily on AI Infrastructure
Building large AI systems requires significant computing power. Tencent has been increasing its spending on AI infrastructure as the competition becomes more intense.
According to The Edge Malaysia, Tencent increased capital spending on AI projects and computing by 176% to 52.8 billion yuan, or about $7.9 billion, in the quarter ended June. (The Edge Malaysia)
That investment gives important context to the Tencent Hy4 AI Model launch. Developing an advanced model is only one part of the AI race. Companies also need chips, data centers, software infrastructure and enough computing capacity to train and serve models to users.
Tencent has an advantage that many smaller AI companies do not have: a large existing business ecosystem. Its social platforms, gaming operations and enterprise products can potentially support continued AI investment.
At the same time, heavy spending does not guarantee success. AI models need to deliver useful results at a reasonable cost, and users can quickly switch between competing services if another product works better.
Hy4 vs GLM and Kimi: What the Numbers Tell Us
The available Tencent evaluation provides an interesting snapshot of the competition.
| Model | Tencent Internal Average Score | Evaluation |
|---|---|---|
| Hy4 Preview | 2.99/4.00 | Tencent’s highest result |
| Kimi K3 | 2.94/4.00 | Close competitor |
| GLM-5.3 | 2.92/4.00 | Close competitor |
The gap between the three models is narrow. That means the Tencent Hy4 AI Model should not be described as having completely dominated its rivals. A more accurate description is that Tencent has produced a competitive model that performed slightly better in its own blind engineering evaluation.
Independent benchmarks will be particularly important in the coming months. They can help show whether Hy4 maintains its performance across coding, reasoning, mathematics, research, writing and long-context tasks.
What Could Come Next for Tencent Hy4?
Tencent has already indicated that the current release is a preview. The company says future versions will improve through additional pre-training and post-training. (Tencent)
The Tencent Hy4 AI Model also has known limitations. Tencent’s technical documentation says the early version can sometimes spend more time than necessary reasoning through complex tasks and may over-check its own work. (GitHub)
That admission is useful because it shows that the model is still being developed. A preview release does not represent the final capability of the Hy4 family.
The next stage will likely focus on making the model faster, more accurate and more efficient. Tencent will also need to keep improving its integration with products such as WorkBuddy and CodeBuddy so that model improvements translate into better user experiences.
Why This Launch Matters for the Global AI Market
The Tencent Hy4 AI Model is another sign that the AI competition is becoming broader. The market is no longer controlled by a small number of companies from one region.
Chinese technology companies are developing increasingly capable models, while global developers are gaining access to more open options. This competition can benefit users because companies have stronger reasons to improve performance and offer better tools.
There is also an important business angle. AI is moving from simple chat applications toward tools that can complete real tasks. Coding, office work, research, data analysis and creative development are becoming major areas of competition.
Tencent’s strategy fits this change. Rather than focusing only on chatbot conversations, the company is connecting Hy4 with products that people can use for everyday work.
Final Analysis
The Tencent Hy4 AI Model is an important new entry in China’s AI competition, but it is still too early to call it the clear winner over Moonshot or Z.AI. Tencent’s internal testing shows a narrow advantage over GLM-5.3 and Kimi K3, while the model’s 770 billion parameters, 49 billion active parameters and more than one-million-token context window give it an impressive technical foundation. (Tencent)
The bigger story is the growing choice available to AI users. Tencent, Moonshot, Z.AI and other companies are developing different approaches, and each model may perform better for different users and tasks.
For now, the Tencent Hy4 AI Model looks like a serious competitor rather than a guaranteed market leader. Its open-source release, large context window and focus on practical work make it worth watching. However, independent testing, real-world adoption and long-term reliability will ultimately tell us how strong Hy4 really is.
As the AI industry continues to develop, healthy competition can give developers and users more options. The most valuable model will not necessarily be the one with the biggest parameter count. Instead, the winners are likely to be the systems that are useful, reliable, affordable and easy for people to use.
Sources
- Tencent — Official Hy4 Preview Announcement: Tencent official Hy4 preview announcement
- Tencent Hunyuan Hy4 Preview — Technical Documentation: Tencent Hy4 Preview on GitHub
- Hugging Face — Tencent Hy4 Preview Model Card: Tencent Hy4 Preview on Hugging Face
- India Today — Tencent Hy4 Launch Report: India Today report on Tencent Hy4
- The Edge Malaysia — Tencent AI Model and Investment Report: The Edge Malaysia report
PayPal Takeover Collapse: Shocking $53 Billion Deal Falls Apart As Shares Tumble
