Tencent is developing Hy4, the next-generation large language model in its Tencent Hy family, as the company looks to scale up its artificial intelligence capabilities following the launch of Hy3. Tencent has confirmed that Hy4 is currently being trained, will have a larger parameter scale than Hy3 and is planned for release during 2026. However, the company has not yet publicly disclosed Hy4’s final parameter count, architecture, context length, benchmarks, pricing or model license.
The development follows the rapid evolution of Tencent’s Hy3 generation. Hy3 Preview was released as an open-weight model in April 2026 under the commercially permissive Apache 2.0 license, while the full Hy3 model was officially launched in July with 295 billion total parameters, 21 billion active parameters and support for a context window of up to 256K tokens. Tencent says the experience gained from Hy3 is being used to develop larger models with stronger reasoning, coding and agent capabilities.
Tencent Confirms Larger Hy4 Model
Tencent’s latest disclosures indicate that the company’s AI team has moved toward a larger model after Hy3. The company said it is training Hy4 with a parameter scale larger than Hy3 and plans to release it within 2026.
The announcement is significant because Hy3 already represents a major step up in Tencent’s open-model strategy. Rather than simply increasing model size, Tencent has emphasized reinforcement learning, better datasets and real-world product feedback as key components of the next stage of development.
What Tencent Has Confirmed About Hy4
| Metric | Hy4 Status |
|---|---|
| Developer | Tencent Hy |
| Generation | Hy4 |
| Development status | Currently being trained |
| Planned release | 2026 |
| Parameter scale | Larger than Hy3 |
| Exact parameters | Not disclosed |
| Architecture | Not disclosed |
| Context length | Not disclosed |
| Benchmarks | Not yet published |
| API pricing | Not disclosed |
| Open weights | Not yet officially confirmed |
| License | Not yet officially confirmed |
The distinction between confirmed development and formal model release is important. Current public information does not support claims that Hy4 Preview has already been released or that Tencent has officially assigned it an Apache 2.0 license.
Hy3 Provides The Foundation For Hy4
Hy4 follows Hy3, which Tencent has positioned as a major upgrade in reasoning, coding and agentic capabilities.
Hy3 Preview was released on April 23, 2026, with its model weights made available through Hugging Face, ModelScope and GitCode. The model was released under Apache 2.0, allowing developers to use, modify and deploy it under the license’s terms.
The preview model uses a Mixture-of-Experts architecture with 295 billion total parameters and 21 billion active parameters.
Hy3 Preview Technical Profile
| Specification | Hy3 Preview |
|---|---|
| Total parameters | 295 billion |
| Active parameters | 21 billion |
| MTP parameters | 3.8 billion |
| Architecture | Mixture of Experts |
| Layers | 80 + MTP layer |
| Attention | GQA |
| KV heads | 8 |
| License | Apache 2.0 |
| Weight availability | Open weight |
| Initial release | April 23, 2026 |
Tencent described Hy3 Preview as its strongest model at the time and highlighted improvements in complex reasoning, instruction following, context learning, coding and agent tasks.
Tencent Has Already Released Full Hy3
The progression from Hy3 Preview to the full Hy3 model also provides clues about Tencent’s development strategy.
Tencent officially launched Hy3 on July 6, describing it as a more stable and cost-efficient version of the model developed from the preview release. The company said Hy3 delivered intelligence comparable to flagship models with two to five times its parameter scale.
Hy3 also expanded its integration across Tencent’s own products, including WorkBuddy, CodeBuddy, Yuanbao, Marvis and ima.
Tencent Hy Model Timeline
April 2026
│
└── Hy3 Preview released
│
├── 295B total parameters
├── 21B active parameters
└── Apache 2.0
│
▼
July 2026
│
└── Full Hy3 officially launched
│
├── Stronger reasoning
├── Better agent capabilities
└── Broader Tencent product integration
│
▼
2026
│
└── Hy4 under training
│
├── Larger than Hy3
├── More reinforcement learning
└── Planned release during 2026
Tencent said it completed the development cycle from rebuilding its AI infrastructure to launching Hy3 Preview and subsequently improving the model in less than six months.
Reinforcement Learning Becomes A Bigger Priority
Tencent’s development plans indicate that Hy4 is not simply about increasing the number of parameters.
The company has said its larger-model effort involves building bigger and better datasets and scaling reinforcement learning to improve contextual understanding, agent capabilities and coding. Tencent is also working with its product teams to identify high-value use cases and optimize data selection and reinforcement learning around them.
This approach reflects the broader shift in AI development toward post-training and reasoning improvements.
Larger models can provide greater capacity, but companies increasingly use reinforcement learning and real-world feedback to make models more effective at multi-step tasks, software development and autonomous agent workflows.
Areas Tencent Is Targeting
| Development Area | Hy4 Direction |
|---|---|
| Model scale | Larger than Hy3 |
| Reinforcement learning | Increased |
| Training data | Larger and improved datasets |
| Coding | Stronger agentic coding capabilities |
| Contextual understanding | Continued improvement |
| General intelligence | Broader capability |
| Product feedback | Used for iterative development |
| Multimodal capabilities | Continued expansion |
Tencent has also said that future Hy models will continue to improve multimodal capabilities while using real product usage and domain feedback to identify weaknesses and accelerate model iteration.
Open-Weight Strategy Has Given Tencent Global Reach
Tencent’s approach to Hy3 has helped the company reach developers outside its own ecosystem.
The Apache 2.0 license used for Hy3 allows developers to download and deploy the model and provides broad freedom for adaptation and commercial use subject to the license terms. Tencent has also made Hy3 available through global developer platforms and open-source communities.
OpenRouter has become an important distribution channel. Tencent said Hy3 became the most-used model on OpenRouter by token usage from April 28, shortly after the preview release.
Hy3’s Global Distribution
| Platform / Channel | Hy3 Availability |
|---|---|
| Hugging Face | Yes |
| ModelScope | Yes |
| GitCode | Yes |
| OpenRouter | Yes |
| Tencent Cloud TokenHub | Yes |
| Other global developer platforms | Progressive integration |
| License | Apache 2.0 |
This distribution model could become strategically important if Tencent applies a similar approach to future generations, although the company has not yet confirmed that Hy4 will use the same license or open-weight model structure.
Agentic AI Is A Major Focus
One of Tencent’s clearest priorities for the Hy family is agentic AI.
Hy3 has been integrated into products that perform multi-step tasks rather than simply answering questions. Tencent’s Yuanbao, for example, has an Agent function capable of handling complex tasks and generating files in formats including PowerPoint, Word, Excel, PDF and HTML.
Hy3 is also being used in productivity and software-development environments, giving Tencent access to real-world data on how developers and users interact with AI systems.
That feedback can potentially help Tencent train Hy4 for more reliable planning, tool use, coding and task execution.
Tencent Faces Strong Competition
Hy4 is being developed in an increasingly competitive open-model market.
Chinese technology companies are investing heavily in large language models, while global developers have access to open-weight systems from multiple organizations. The competition is increasingly centered on reasoning quality, coding performance, inference costs, context length and agent capabilities rather than parameter count alone.
Tencent’s advantage is its ability to combine model development with a large ecosystem of consumer and enterprise products.
The company can test models through real-world applications such as productivity assistants, gaming, communications and cloud services before making improvements to subsequent generations.
The License Question Remains Open
The Apache 2.0 license is one of the most important features of Hy3’s open-model strategy, but it should not automatically be attributed to Hy4.
As of the latest available information, Tencent has confirmed that Hy4 is being trained and is expected to launch in 2026, but has not officially published its model card, weights or license.
That means developers should wait for an official Hy4 announcement before assuming that its weights will be freely downloadable or that it will carry the same Apache 2.0 terms as Hy3.
The Bigger Picture
Tencent’s Hy4 development shows that the company is accelerating its AI model roadmap after the rapid progression of Hy3. The company is training a larger model while increasing its focus on reinforcement learning, better datasets, agentic coding and multimodal capabilities. Hy3’s open-weight Apache 2.0 release has also demonstrated Tencent’s willingness to compete for global developer adoption rather than keeping its strongest models entirely behind proprietary APIs.
The key question for Hy4 will be whether Tencent continues that open-model strategy. If Hy4 eventually arrives with open weights and a commercially permissive license, it could become another major option for developers building AI agents, coding systems and enterprise applications. For now, however, the confirmed story is that Hy4 is under development and planned for release in 2026; claims of an already released “Hy4 Preview” under Apache 2.0 remain unverified by the official sources reviewed.
Looking Ahead
Tencent is expected to provide more technical details about Hy4 as development progresses, including its final parameter scale, architecture, benchmark performance, deployment options and licensing model. The company has already confirmed that the next generation will be larger than Hy3 and will build on the infrastructure, training techniques and real-world feedback accumulated during the Hy3 development cycle.
For the open-source AI community, the eventual licensing decision could be particularly significant. Hy3’s Apache 2.0 release helped Tencent establish a presence among global developers, while the company’s large consumer and enterprise ecosystem provides a substantial testing environment for future models. If Hy4 combines a larger scale with stronger reasoning and agent capabilities while retaining an open-weight approach, it could substantially increase Tencent’s influence in the rapidly expanding global AI model market.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
