Tencent Releases Hy4, a 770B-Parameter Open-Weight Model

Tencent launched a new AI model called Hy4 on November 4, 2026. This model is the first in the Hy line capable of understanding images and operating computers. Hy4 is a mixture-of-experts model with a total of 770 billion parameters. This model is one of the strongest open-weight models available today. The launch marks a significant milestone for Tencent in the global AI competition which is increasingly fierce.

Compared with Hy4 preview, this new model scales both pre-training and post-training significantly. Vision capability is added for the first time in this model line. Hy4 also supports computer use, allowing the model to work through command-line interfaces and graphical user interfaces. The model can understand fine-grained details in images and analyze complex charts. With computer use, Hy4 can perform multi-step tasks on a computer. This model is not just answering questions, but a model that can look at a screen and complete work on it.

Access to Hy4 is already available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima, and WeCom assistant Dayuan. API is also available through OpenRouter and other gateways. However, model weights cannot be downloaded yet. Tencent plans to open-source the weights, but has not given a specific date for the release. This policy has sparked discussion in the research community about the true definition of open-weight.

Technical Architecture of Hy4 Model

Hy4 uses a mixture-of-experts (MoE) architecture with a total of 770 billion parameters. The MoE architecture allows the model to activate only a small subset of parameters for each inference. This makes the model more efficient compared to dense models of the same size. This design follows the industry trend popularized by models like Mixtral and DeepSeek MoE. Vision capability in Hy4 is implemented through a vision encoder integrated with the model backbone. The vision encoder processes images into vector representations that are then fused into the language model. This approach enables deep multimodal understanding without retraining the entire model. Computer use capability is built on top of the vision foundation by adding an action head for interface control.

Deeper Dive: Hy4's Position in the Open-Weight Model Market

Hy4 breaks into the top tier among open-weight models available today. Tencent claims that this is its entry into the top tier of open-source models. However, there is still a significant gap between Hy4 and top proprietary models available today. Benchmark comparisons show specific strengths and weaknesses of this model.

Here is a comparison of several benchmark results between top models:

Hy4 leads on APEX-Agents, but loses big on Terminal-Bench. This shows the model is still in development and has not yet reached peak capability. Open weights are also not released yet, so the community cannot independently verify benchmark results. Transparency is a critical issue for enterprise adoption which requires independent verification.

API pricing for Hy4 remains the same as the preview version. Input costs 6 yuan per million tokens. Output 18 yuan per million tokens. Cache hit 0.3 yuan per million tokens. Tencent will also release Hy4 flash soon, a more cost-efficient version with faster inference speed. This pricing strategy positions Hy4 as a competitive alternative for developers in China.

Implications for developers are also interesting. Open-weight models give developers the opportunity to try advanced models without expensive licensing fees. However, the lack of open weights remains a barrier for broader local use. For China's AI ecosystem, the Hy4 release shows that Chinese companies continue to survive in the global AI model competition. Even though there is still a gap with top proprietary models from the United States, Hy4 proves that open-weight models from China can compete at the top level and drive continuous innovation in the regional market.

Impact on AI Ecosystem and Developers

The arrival of models like Hy4 has quite an interesting impact on the AI ecosystem as a whole. First, open-weight models give local developers the opportunity to try advanced models without expensive licensing fees. Second, API availability makes integration into local applications easier. Third, competition between open-weight models from China and proprietary models from the US can drive overall prices down.

However, there are several challenges that need to be addressed to maximize the benefits of this model. First, the lack of open weights prevents developers from doing fine-tuning locally. Second, dependence on foreign APIs can be a problem if there are network restrictions or cross-border regulations. Third, documentation and support in local languages are still very limited. Fourth, the absence of open weight release delays independent verification by the research community.

Therefore, Hy4 is a positive step for the AI ecosystem, but more attention is needed from both government and private sector to maximize its benefits. Supportive regulation and compute infrastructure investment will determine how far this model can contribute to national technological sovereignty. Collaboration between academia, industry, and government becomes the key to domestic AI ecosystem success that is sustainable. Inclusivity for the digital future must be maintained as a shared priority.

Tencent's Strategy in Global AI Competition

Tencent's strategy with Hy4 reflects a different approach from its main competitors. Instead of chasing closed proprietary models, Tencent focuses on the open-weight ecosystem hoping to build a strong developer community. This approach aligns with Meta's strategy with Llama and Mistral AI with their models. However, the delay in open weight release raises doubts among industry observers. Many question whether Tencent is truly committed to open-source philosophy or just using the term for marketing purposes. The answer to this question will become clear in the coming months when Tencent announces a concrete schedule for weight release. Long-term transparency will determine community trust in the Tencent ecosystem. The sustainability of its adoption in the competitive global market depends on this consistency.

Comparison With Other Open-Weight Models

Besides Hy4, there are several other open-weight models that also attract attention from the AI community. Meta's Llama 3.5 405B remains the benchmark for large-scale open-weight models. DeepSeek V3 with its 671B parameter MoE architecture offers competitive performance with lower inference costs. Alibaba's Qwen 2.5 72B shows that medium-sized models remain relevant for many use cases.

Hy4 is different because it has far more parameters (770B) compared to other open-weight models from China. This parameter size advantage should provide greater reasoning capacity. However, MoE efficiency means only a subset of parameters is active per inference. Realizing the full potential of this model depends heavily on training quality and data used. This competition drives continuous innovation across the entire AI industry. End users directly benefit through more diverse choices.

Future Prospects for Hy4 Model

The future of Hy4 depends heavily on the execution of the promised open weight release by Tencent. If weights are fully released with a permissive license, Hy4 could become the foundation for thousands of derivative projects and domain-specific fine-tunes. The research community can also conduct independent security and bias audits to increase public trust. This milestone will define the trajectory of open AI development in the region for years to come.

Tencent's roadmap includes Hy4 flash release for fast inference and distilled variants for edge deployment. Multilingual support will be expanded including Indonesian for the Southeast Asian market. Deeper integration with the Tencent Cloud ecosystem will facilitate large-scale enterprise adoption. Digital transformation across various strategic industrial sectors will be accelerated significantly.

Key Terms Glossary

Key terms in this article include several fundamental technical concepts. Hy4 is Tencent's 770 billion parameter open-weight AI model. MoE or Mixture-of-Experts is a model architecture that divides tasks among several experts for computational efficiency. Vision refers to the model capability to understand images and process visual information. Computer Use is the model capability to operate computer interfaces like a human. Open Weight means model weights available to the public for free. OpenRouter is a gateway to access various AI models through a unified API.