Security · November 7, 2026
Tencent releases Hy4 AI model with restricted cybersecurity training data
Tencent has released its latest artificial intelligence model, Hy4, which is designed to limit its ability to learn offensive cyberattack techniques. The company stated that the model was trained primarily on defensive data, such as software vulnerability detection and patching, to reduce the risk of malicious use. Tencent declared that Hy4 now performs on par with the top-tier open-weight AI models in China.
The model was announced on the fourth of the month as a successor to the Hy4 preview released in August. It features an expanded scale for both pre-training and post-training. The official model weights have not yet been made public, but Tencent plans to distribute them soon alongside a system card containing safety evaluation results.
Hy4 is built on a mixture-of-experts architecture with a total of 770 billion parameters. During inference, it activates approximately 49 billion parameters. The model supports a context window of up to one million tokens. It introduces image understanding capabilities for the first time, enabling the analysis of complex charts and mathematical visual reasoning. The model also adds computer operation features, allowing it to handle both command-line and graphical user interfaces.
Training included specialized software use, such as Blender and CAD. The model was enhanced to plan and execute long-duration tasks across multiple applications. Performance in frontend development and game creation was also improved. Tencent emphasized that Hy4 achieves comparable performance to the latest models from Z.ai, Moonshot AI, and DeepSeek in coding, agent, and browser usage evaluations.
In an evaluation involving 213 financial experts and 579 practical tasks, Hy4 scored 3.04 points. This result surpassed DeepSeek V4.1 Flash, which scored 2.97 points, and Kimi K3.1 Pro and GLM-5.3, which each scored 2.68 points. In a separate work evaluation using the WorkBuddy tool, Hy4 ranked first among Chinese models in nine of 12 assessment categories. The overall task pass rate reached 96.15 percent. Internal evaluations showed that the hallucination rate dropped from 12.95 percent in the preview version to 7.2 percent.
External evaluations, however, showed that Hy4 still lags behind the highest-performing models in China. In a test released by Vals AI, Hy4 achieved an overall accuracy of 59.35 percent, ranking 13th out of 33 models and fourth among open-weight models. It trailed Kimi K3.1 Pro, DeepSeek V4.1 Flash, and GLM-5.3. The model did improve upon the previous preview, which scored 53.21 percent. In the Terminal Bench 4.0 coding agent evaluation, Hy4 ranked second among open-weight models with a score of 39.39 percent.
Tencent acknowledged a gap between Hy4 and top-tier closed models. Significant differences remain in research-level mathematics, long-duration coding, model self-improvement, and advanced visual question answering. In the cybersecurity domain, training data was restricted to defensive purposes. Materials that could be directly used for system intrusion or vulnerability exploitation were excluded. Reverse engineering and simulated penetration testing data were used only within the scope necessary for defensive analysis.
Security evaluations by Tencent's Xuanwu Lab and Keen Security Lab found no significant risks related to offensive capabilities. The labs assessed cryptographic analysis, binary deobfuscation, and vulnerability exploitation. Biological safety evaluations indicated that while the model possesses basic biology knowledge, its ability to design complex experiments does not match that of top-tier models.
The release comes amid heightened concerns about AI-enabled cyberattacks. Last month, Tencent disclosed that agents testing external AI models exhibited unintended web browsing behavior. The company stated that this behavior did not originate from its own Hunyuan model and that no evidence of intentional attacks or organized activity was found.
Hy4 is currently available on Tencent Cloud, CodeBuddy, WorkBuddy, and Yuanbao. API pricing remains the same as the previous preview. WorkBuddy and CodeBuddy will offer a certain amount of free usage until the 15th. Tencent plans to release the official model weights, system card, and a lighter model called Hy4 Flash, which will focus on improved inference speed and cost efficiency.