Tencent Releases Hy4 Multimodal Model With Computer-Use Features and Planned Open Weights

Tencent announced the launch of Tencent Hy4, the latest generation of its flagship artificial intelligence model, introducing visual understanding and computer-use automation while committing to release the underlying model weights and system cards in the coming days.

According to a company announcement published at https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/, Hy4 marks the first entry in the Hy family to incorporate native vision capabilities, enabling multimodal tasks such as chart interpretation, fine-grained image processing, and visual mathematical reasoning. The release reflects an accelerating pattern among major Chinese technology companies expanding access to open-weights models while embedding autonomous task capabilities directly into enterprise software.

Hy4 also introduces computer-use controls, allowing the system to interact directly with command-line prompts and graphical user interfaces. Tencent reported training the model specifically on complex technical software, including Blender and computer-aided design platforms, to improve multi-step planning and long-horizon execution.

The model is being deployed immediately across Tencent's workplace products. In Tencent WorkBuddy, an enterprise productivity application tied to WeCom, Hy4 handles project timelines, deadline monitoring, and meeting preparation. On internal WorkBuddy benchmarks reported by the company, Hy4 recorded an average score of 83.18 and a 96.15 percent pass rate, placing first in nine of twelve monitored mainland Chinese workplace scenarios.

Tencent also targeted professional financial analysis, stating it trained Hy4 on specialized workspaces reflecting investment research workflows. In a vendor-organized blind evaluation involving 213 financial analysts assessing 579 tasks, Hy4 scored 3.04, compared to 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3. Tencent reported that factual retrieval tuning reduced the model's hallucination rate from 12.95 percent in the Hy4 Preview to 7.2 percent, while errors stemming from unsupported assumptions dropped from 23.1 percent to 5.6 percent.

For software and game production, the model includes integrations with Unreal Engine, Unity, and Godot. Tencent stated Hy4 achieved a 19.1 percent improvement in internal GameDev testing and a 9.8 percent gain in GameLogic evaluations compared to its preview build. In scientific domains, the company evaluated the model against TenureTrail, a 123-task benchmark spanning seven disciplines, and open-sourced two specialized foundational models built using Hy4: HyCell for single-cell biology and HyPDE for partial differential equations.

Addressing model safety, Tencent stated that cybersecurity training was confined to defensive tasks such as vulnerability remediation and binary reverse engineering, with offensive exploitation materials excluded from the training set. Internal assessments carried out by Tencent's Xuanwu Lab and Keen Security Lab identified no significant offensive risk across evaluated cryptographic reverse engineering and binary deobfuscation tests. Internal biosafety reviews noted that while Hy4 demonstrated foundational biological knowledge, its capacity for open-ended experimental design remains constrained compared to leading frontier models.

Hy4 is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, and Yuanbao, as well as via third-party API providers including OpenRouter, Hermes, and OpenCode. Tencent maintained API pricing from the preview version at approximately $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens, with a lighter Hy4 Flash variant planned for subsequent release.