Tencent Releases Hy4 AI Model with Restricted Cyberattack Training
November 7, 2026 · 2 min read
Also published in Deutsch, Italiano, Nederlands, Русский, עברית, Português (Brasil), Svenska
Tencent has released Hy4, a new artificial intelligence model designed to minimize the potential for misuse in cyberattacks. The model was trained primarily on data intended for defensive purposes, such as identifying and fixing software vulnerabilities, while excluding attack materials that could be directly used for system intrusions or exploiting vulnerabilities.
Hy4 is based on a Mixture of Experts architecture with 770 billion parameters, activating approximately 49 billion during inference. It supports a context window of up to 1 million tokens and introduces image understanding capabilities for mathematical visual reasoning and complex chart analysis. The model also includes computer operation functions for both graphical user interfaces and command line interfaces, and was trained for professional software such as CAD and Blender.
Tencent stated that Hy4 achieves performance comparable to top open weight models from DeepSeek, Moonshot AI, and Z.ai in evaluations for browser usage, agents, and coding. In a practical task evaluation involving 579 assignments and 213 financial experts, Hy4 scored 3.04 points, surpassing GLM-5.3 and Kimi K3.1 Pro, which both scored 2.68 points, and DeepSeek V4.1 Flash, which scored 2.97 points. In internal tests for the WorkBuddy AI tool, the model ranked first among Chinese models in 9 out of 12 categories with a total task pass rate of 96.15 percent, while hallucination rates dropped to 7.2 percent from 12.95 percent in the preview version.
External evaluations show some gaps remain. A report from Vals AI ranked Hy4 13th out of 33 models with an overall accuracy of 59.35 percent, trailing behind GLM-5.3, DeepSeek V4.1 Flash, and Kimi K3.1 Pro. However, it ranked second among open weight models in the Terminal Bench 4.0 coding agent evaluation with 39.39 percent. Tencent acknowledged significant differences still exist compared to top closed models in areas such as advanced visual question answering, long term coding, and research level mathematics.
Security assessments by Xuanwu Lab and Keen Security Lab found no significant risks regarding attack capabilities. Biological safety tests indicated that while the model possesses basic biological knowledge, it cannot design complex experiments as well as top models. Hy4 is available via Tencent Cloud, Yuanbao, CodeBuddy, and WorkBuddy. The company plans to release the official model weights, a system card, and a lightweight Hy4 Flash model.