IT之家 11 月 4 日消息,腾讯混元今天(4 日)发布最新一代大语言模型 Hy4 正式版。官方表示,随着模型能力的再次全面跃升,Hy 首次跻身最强开源模型之列。Hy4 将于近期开源模型权重,发布模型平台卡片,并积极推进更负责任的开源,重点关注网络和生物安全等风险。
相比于 Hy4 preview,Hy4 显著扩展了预训练和后训练规模,首次引入了视觉能力,并大幅提高了指令遵循、长文理解、长程任务、深度研究等各项综合能力。
Hy4 首次引入视觉能力,可以结合感知、推理和工具调用,开展细粒度图像处理、复杂图表分析和数学视觉推理等任务。Hy4 也首次兼容电脑操作,可以灵活地结合使用命令行和图形界面。腾讯还针对 Blender、CAD 等专业软件和跨应用长程任务开展了专项训练,持续优化专业任务的规划与执行力。
依据介绍,该模型可用于视觉与电脑任务、前端与游戏开发、金融与办公情形、科研与模型研发、长期个人助理等多种情形,为真实生产力而生。
据IT之家此前报道,8 月 28 日,腾讯发布 Hy4 preview,总参数 770 B,激活参数 49 B,上下文长度 1M。
官方当时曾表示,Hy4 preview 在模型尺寸、上下文长度、数据规模上都开展了显著的扩展,预训练和后训练的共同进步带来了智能水平的又一次巨大提高。
当下,Hy4 已在腾讯云和 OpenRouter 上线,并可在 CodeBuddy / WorkBuddy、元宝、ima 等机型上感受,表现显著超过 Hy4 preview 的同时维持相同 API 售价。
官方计划开源模型权重,并已联合 vLLM/SGLang 社区以及 Inferact Inc. 给出开源生态兼容。模型平台卡片也计划届时发布,涵盖更全面的安全评估等细节。
尽管 Hy4 已经是当下最强的开源模型之一,其与顶尖闭源模型仍然存在较大差距,尤其在内部构建的前沿数学、长程开发、模型迭代、视觉问答等领域的高难度测试集上。腾讯混元维度表示,将持续稳健地扩展预训练和后训练规模,努力缩小和顶尖闭源模型的差距。
Hy4 flash 也计划近期发布,将显著提高模型性价比和推理速度,并更加高频地和机型协同迭代,不断提高实用性。
本文来源:https://www.ithome.com/1/022/035.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT Home users Zhixia, Domado, McWings, xxy171070, Alita's Fan, Onlooker_, Yae Miko, and Weiyang for their tips!
IT Home, November 4 — Tencent Hunyuan today (the 4th) released the official version of its latest large language model, Hy4. Officials stated that with another comprehensive leap in model capabilities, Hy has joined the ranks of the strongest open-source models for the first time. Hy4 will open-source its model weights soon, release a model platform card, and actively promote more responsible open-sourcing, focusing on risks such as cybersecurity and biosecurity.
Compared with Hy4 preview, Hy4 has significantly expanded its pre-training and post-training scales, introduced visual capabilities for the first time, and greatly improved comprehensive capabilities including instruction following, long-context understanding, long-horizon tasks, and deep research.
Hy4 introduces visual capabilities for the first time, combining perception, reasoning, and tool calls to carry out tasks such as fine-grained image processing, complex chart analysis, and mathematical visual reasoning. Hy4 also supports computer operations for the first time, allowing flexible combined use of command line and graphical interfaces. Tencent has also conducted specialized training for professional software like Blender and CAD, as well as cross-application long-horizon tasks, continuously optimizing planning and execution capabilities for professional tasks.
According to the introduction, the model can be used in various scenarios such as visual and computer tasks, front-end and game development, finance and office scenarios, scientific research and model development, and long-term personal assistants, built for real-world productivity.
As previously reported by IT Home, on August 28, Tencent released Hy4 preview, with 770 B total parameters, 49 B active parameters, and a 1M context length.
Officials stated at the time that Hy4 preview underwent significant expansions in model size, context length, and data scale, and the joint advancement of pre-training and post-training has brought another massive improvement in intelligence level.
Currently, Hy4 is available on Tencent Cloud and OpenRouter, and can be experienced on platforms such as CodeBuddy / WorkBuddy, Yuanbao, and ima. It performs significantly better than Hy4 preview while maintaining the same API pricing.
The company plans to open-source the model weights and has already partnered with the vLLM/SGLang communities and Inferact Inc. to provide open-source ecosystem compatibility. A model platform card is also planned for release at that time, covering more comprehensive details such as safety evaluations.
Although Hy4 is already one of the strongest open-source models available today, there remains a significant gap compared to top closed-source models, especially on high-difficulty test sets in areas such as internally constructed cutting-edge mathematics, long-horizon development, model iteration, and visual question answering. Tencent Hunyuan Dimension stated that it will continue to steadily expand its pre-training and post-training scales, striving to narrow the gap with top closed-source models.
Hy4 flash is also scheduled for release soon, which will significantly improve the model's cost-performance and inference speed. It will also iterate in synergy with the models more frequently, continuously enhancing practicality.


