Tencent has introduced Hy4, the latest version of its AI model family, following a preliminary release in August. The company attributes its improvements to broader training and work on instruction following, long-context handling and multistep tasks.
Hy4 adds vision and computer-use capabilities, allowing it to interpret images, analyze complex charts and interact with software through commands or graphical interfaces. Tencent says the model can also work with specialist applications such as Blender and CAD tools, as well as assist with programming and video-game development.
A notable choice in Hy4’s development concerns cybersecurity. Tencent says it prioritized defensive information, such as material for identifying and fixing vulnerabilities, while excluding offensive content about exploiting them. The company describes this as a decision about what information to include in training, rather than simply a restriction on how the model responds. It does not, however, establish that the approach prevents malicious use.
Tencent says Hy4 still includes knowledge of reverse engineering and penetration testing for analytical purposes. The company’s stated distinction is between using those disciplines to assess security and training on material intended to exploit weaknesses offensively.
In a comparison published by Tencent, Hy4 scored 44.4 on Terminal-Bench 4.0, below GPT-6.1 Sol at 55.5 and Kimi K3.1 Pro at 66.8. In a separate evaluation of 579 financial tasks completed by 213 professionals, Tencent reports that Hy4 scored 3.04, narrowly ahead of DeepSeek V4.1 Flash at 2.97. The Arena agent ranking dated November 3 placed Hy4 14th, behind Gemini 4 Carbon in second place and GPT-6.1 Sol in sixth. These assessments cover different tasks and use different methods, so they do not provide a single overall ranking.
Tencent says specialists from its Xuanwu Lab and Keen Security Lab tested Hy4 on areas including cryptographic reverse engineering, binary-file deobfuscation and vulnerability exploitation. The company reports that they found no significant issues involving offensive capabilities in the areas assessed. Those findings come from Tencent; the company says model weights and technical security documentation will be released later for further evaluation.
Tencent is not the only company to explore removing cybersecurity information during model development. The source account notes that Moonshot AI applied a similar approach to Kimi K3.1 Pro, while the UK’s AI Security Institute has examined filtering potentially dangerous biological information in experimental models. Those examples do not establish how well the approach works for Hy4.
Sources: Xataka.