腾讯发布 Hy4 前沿大模型，首次加入视觉理解与电脑操作能力
Tencent releases Hy4, powerful frontier LLM introducing visual understanding and computer use - SiliconANGLE
腾讯发布大语言模型 Hy4，首次加入电脑操作能力，结合视觉理解、推理与工具调用实现长程智能体表现。该模型与 8 月 28 日的 Hy4 Preview 同为 7700 亿参数、490 亿激活参数和 100 万 token 上下文窗口，在 GameDev bench 和 GameLogic 基准上分别提升 19.1% 和 9.8%，但基线数值未公开。模型已在腾讯云 TokenHub、CodeBuddy、元宝等平台上线，也可通过 OpenRouter 等第三方渠道访问，定价维持每百万输入 token 约 90 美分、输出 2.70 美元、缓存输入 4.5 美分，Flash 版本与模型权重将在近日发布。
正文
Tencent Holdings Ltd. today introduced Hy4, a new large language model that adds computer-use capabilities for the first time, using advanced visual understanding, reasoning and tool-calling to deliver long-horizon agentic performance.
The company said the model marks a significant leap in performance over its Hy4 Preview, which arrived Aug. 28, weighing in at 770 billion parameters with 49 billion active parameters. Both provide a 1 million-token context window, putting it in the same range as current open-weights frontier models from European and Chinese developers, including Mistral AI SAS, Beijing Moonshot Technology Co. Ltd. and Z.AI Co. Ltd.
The company said it substantially scaled up both pre-training and post-training, reporting gains over the preview version in instruction following, long-context understanding, agentic long-horizon tasks and deep research. Tencent also identified excessive reasoning and over-verification during the previous version’s “thinking” processes.
The new model focuses heavily on real-world productivity for workplace collaboration, human experience and professional work. The company said it optimized long-context learning capabilities to understand complex information, including web development, document summarization, visual creative tasks, document presentation, and more.
The company added that it also enhanced its capability for front-end and game development, high-complexity tasks that include intricate business logic, state management and orchestration. Some models struggle with producing stateful, elegant and beautiful user interfaces that are visually polished and also fully functional.
It is not uncommon for professional game developers, industry engineers and amateur vibe coders to use AI models as part of their work. However, a recent spate has had coders begin to reproduce, reverse-engineer and clone older and semi-recent video games. That prompted Anthropic PBC to block Claude Code and its underlying Opus 5.5 model from assisting in these tasks.
The company said it integrated deeply with major engines such as Unreal Engine, Unity and Godot, giving it the capability to support complex game logic and multiplayer implementations. Compared wit the preview, Tencent said, Hy4 improved by 19.1 percent in the GameDev bench and 9.8 percent in GameLogic benchmarks, although the actual baseline numbers remain internal and weren’t released.
To advance scientific reasoning, Tencent advanced the next stage of Recursive Self-Improvement to enhance research and development capabilities. Using the Hyra 2.0 harness, an autonomous AI research agent system, the model optimized the pre-training infrastructure for smaller models sharing the same architecture.
The model is currently available on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI Assistant, Ima and more. It can also be accessed via third-party sources such as OpenRouter, Hermes and OpenCode. The company added that, despite its significant enhancements over Hy4 Preview, it kept costs low, maintaining the same pricing at around 90 cents per million input tokens, $2.70 per million output tokens and 4.5 cents per million cached input tokens.
The Flash version of the model is scheduled for release soon to deliver greater cost efficiency for users. The company also said it intends to release the model weights and system cards in coming days.
Image: Tencent