Tencent has introduced Hy4, its latest-generation artificial intelligence model, with a focus on improving AI capabilities while limiting the risks associated with offensive cybersecurity applications. Announced on November 4, 2026, the model has been trained using selected defensive cybersecurity data, including information related to proof-of-concept generation and vulnerability remediation. Tencent says it excluded material involving exploit development, reverse engineering and penetration testing to reduce the possibility of the model being used to support cyberattacks.
The release reflects a growing debate across the AI industry about how developers should balance open access with safety. As AI models become more capable of analysing software and performing complex technical tasks, they could help security teams identify weaknesses faster. However, similar capabilities could also be misused. Tencent says it plans to release Hy4’s model weights and system card, making its safety approach and model capabilities available for further examination.
Why Tencent Is Restricting Hy4’s Cybersecurity Training
The main distinction between Hy4 and many general-purpose AI models is how Tencent has approached cybersecurity training.
Rather than exposing the model to a broad collection of offensive and defensive security material, Tencent says it deliberately limited the cybersecurity content used during training.
The company included defensive information related to identifying software vulnerabilities, generating proof-of-concept demonstrations and developing fixes. It excluded material covering:
- Developing tools for unauthorised hacking.
- Exploiting vulnerabilities to break into systems.
- Reverse engineering software for offensive purposes.
- Penetration testing techniques associated with offensive operations.
The objective is to help developers and security researchers understand and address software weaknesses without significantly improving the model’s ability to carry out attacks.
Tencent also says internal security specialists evaluated Hy4’s capabilities in areas such as cryptographic analysis, reverse engineering, binary obfuscation and vulnerability exploitation.
According to the company, these assessments did not identify particularly elevated offensive cybersecurity capabilities.
However, these are Tencent’s own testing findings. Independent evaluations will be important for determining how effectively the model’s training restrictions translate into real-world safety.
Hy4 Arrives Amid Growing Concerns About AI-Powered Cyberattacks
The launch comes at a time when governments, technology companies and cybersecurity researchers are increasingly concerned about the potential misuse of advanced AI systems.
AI can assist attackers with tasks such as researching software weaknesses, understanding unfamiliar code and automating parts of technical investigations.
The same technology can help defenders analyse suspicious activity, identify vulnerabilities and develop security patches.
This dual-use nature makes cybersecurity one of the most challenging areas for AI developers.
Tencent’s decision to restrict Hy4’s offensive cybersecurity training is part of a broader industry shift towards more controlled model development.
Moonshot AI, another Chinese AI developer, also limited the offensive cybersecurity capabilities of its latest model in October 2026.
The developments suggest that AI companies are increasingly considering safety during model training and release, rather than relying entirely on restrictions applied after a model has been developed.
Hy4 Offers More Than Cybersecurity Capabilities
Although cybersecurity safety is a major part of the launch, Hy4 is designed as a general-purpose productivity model.
Tencent says it introduces visual understanding for the first time in the Hy4 series, alongside improvements in reasoning, instruction following, long-context comprehension and complex agent-based tasks.
Its capabilities include:
| Capability | What it enables |
|---|---|
| Visual understanding | Analysing images, charts and visual information |
| Computer use | Interacting with command-line and graphical interfaces |
| Long-context reasoning | Understanding large amounts of information |
| Coding | Supporting software development and debugging |
| Financial analysis | Assisting with research and quantitative workflows |
| Agentic tasks | Planning and executing multi-step activities |
| Professional software | Working with tools such as Blender and CAD applications |
One particularly important addition is computer-use functionality.
Hy4 can interact with graphical user interfaces and command-line environments to complete tasks. Tencent says it has also trained the model for professional applications such as Blender and computer-aided design software.
This could make the model useful for developers, designers and professionals who need AI assistance across multiple applications.
Tencent Reports Strong Results in Financial Analysis
Tencent has also highlighted Hy4’s capabilities in financial research and analysis.
The company worked with Tencent Financial Technology to evaluate the model using 579 real-world financial tasks. The assessment involved 213 financial professionals, including 153 experts from external financial institutions.
Approximately 70% of the participants had more than five years of professional experience.
In the evaluation, Hy4 achieved an overall score of 3.04, compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3.
| AI model | Evaluation score |
|---|---|
| Tencent Hy4 | 3.04 |
| DeepSeek-V4.1-Flash-1017 | 2.97 |
| Kimi K3.1 Pro | 2.68 |
| GLM-5.3 | 2.68 |
These results come from an evaluation organised by Tencent and its financial technology division. They should not be treated as an independent industry-wide ranking.
Tencent says Hy4 also improved by 12.03 percentage points over Hy4 Preview in the internal Investment Research Agent Benchmark conducted with financial AI company Rabyte.
The results suggest that Tencent is positioning Hy4 for specialised professional applications, including securities research, fund analysis and quantitative investment workflows.
Tencent Plans to Open-Source Hy4’s Model Weights
Tencent says Hy4’s model weights and system card will be released in the coming days.
The company is following an open-weight approach that allows developers to access the model parameters and build applications around the system.
Hy4 is also being integrated into Tencent’s existing products and developer platforms, including WorkBuddy, CodeBuddy, Yuanbao and Tencent Cloud TokenHub. Its API is being made available through platforms including OpenRouter, Hermes and OpenCode.
The model is designed to support a range of applications, from office productivity and software development to financial analysis and scientific research.
Tencent has also maintained the same published API pricing as Hy4 Preview: 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan per million cached input tokens.
The company says a faster and lower-cost Hy4 Flash version is also planned.
What Hy4 Means for the Open-Source AI Industry
Hy4 highlights an important challenge facing companies developing open-weight AI models.
Making model weights publicly available can encourage innovation, independent research and wider adoption. However, releasing increasingly capable models also makes it more difficult for developers to control how their technology is used.
Cybersecurity is particularly sensitive because the same technical knowledge can support both defensive research and malicious activity.
Tencent’s approach attempts to address this challenge by limiting offensive cybersecurity training before releasing the model.
However, removing certain training material does not guarantee that a model cannot produce harmful outputs. AI systems can sometimes develop capabilities through related knowledge, and independent researchers will need to examine Hy4’s actual behaviour after its weights become available.
The Bigger Picture
The release of Hy4 reflects the growing importance of responsible open-weight AI development in China and internationally.
Companies such as Tencent, Moonshot AI, Alibaba and other AI developers are competing to produce increasingly capable models while addressing concerns about cybersecurity, biological risks and autonomous systems.
The industry is moving towards a more complicated approach to openness, where developers must consider not only model performance and accessibility but also the risks associated with releasing powerful capabilities.
For cybersecurity professionals, defensive AI models could become valuable tools for identifying weaknesses and improving software security. For AI developers, Hy4 provides another example of how safety restrictions can be incorporated into the training and release process.
Looking Ahead
The next important milestone will be the public availability of Hy4’s model weights and system card. Independent security researchers will be able to examine the model’s capabilities, test its safeguards and assess whether Tencent’s training restrictions meaningfully reduce offensive cybersecurity risks. The results will help determine whether the company’s approach can serve as a practical model for other developers releasing powerful open-weight systems.
Tencent is also expected to expand Hy4 across its productivity, financial and developer products, while preparing the lower-cost Hy4 Flash version. Its long-term success will depend on whether it can combine competitive AI performance, practical enterprise applications and credible safety controls as competition in the open-weight AI market intensifies.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
