Tencent Hunyuan launches Hy4 with vision and computer use
Tencent Hunyuan launched the 770B-parameter Hy4 model on November 4 and made it selectable in the Yuanbao mobile app, PC client and web interface. According to Tencent, Hy4 adds native visual understanding, computer operation across interfaces and applications, and training for long-horizon agent tasks. Lei Technology, whose article was republished by 36Kr with authorization, conducted hands-on tests that included building an interactive webpage, analyzing an earnings-report screenshot and checking visa-document requirements across WeChat and local files. The report also cites Tencent’s internal testing and crowdsourced evaluations for its claimed reductions in hallucinations and financial-task performance; it provides no independent benchmarks for those claims.
Hy4 combines visual understanding and cross-application computer operation with immediate access through Yuanbao. Lei Technology’s hands-on tests showed the model handling interactive webpage creation, financial-document analysis and cross-application visa-document checks, while the report also described WorkBuddy’s user-confirmation mechanism for sensitive operations.
11月4日，腾讯混元正式推出新一代大语言模型Hy4，并且旗下AI助手元宝也同步完成了能力接入。现在，在元宝App、PC客户端、网页端，用户都能把模型切换到Hy4使用。
（图源：元宝）
据官方介绍，随着多项底层能力的升级，Hy4首次进入了最强开源模型梯队。参数方面，Hy4从Hy3系列的不到300B升级到了770B，但在雷科技（ID：leitech）看来，这次新版本的核心看点在参数之外，它终于补齐了 AI Agent落地的关键能力：视觉理解与电脑操作（Computer Use）。
不出意外的话，Hy4此次加入的新特性，特别是针对长程智能体任务的升级，可能会改变我们对开源模型工作流的预期。现在，我们将结合实际的测试体验，看看它的表现如何。