腾讯发布 Hy4,支持视觉与电脑操作,居中国第二

GPT-Image-2,提示词由 THE DECODER 提供
腾讯发布了其 AI 模型 Hy4,该模型能够分析图像并自主操作软件。模型权重将在稍后发布。
该模型针对编程、金融和办公工作进行了专门优化。腾讯表示,它已经使用 Hy4 来改进其他模型的训练流程。
在独立测试中,Hy4 在中国模型中排名第二,但在高难度数学和编程任务上远远落后于 Claude Opus 5.5 等顶尖模型。
腾讯发布了其 AI 模型 Hy4。它首次能够理解图像,并能自主操作软件。该公司将其列为最好的公开可用模型之一,不过模型权重要到之后才会发布。
在预览版发布约两个月后,腾讯发布了正式版语言模型 Hy4。据腾讯称,该模型训练所用的数据量远超其前代,并“首次”跻身最强开放权重模型之列。在权重发布(具体日期仍未公布)之前,Hy4 只能通过腾讯云、OpenRouter 以及元宝(Yuanbao)聊天机器人等腾讯应用使用。
最大的新功能是图像理解。Hy4 应当能识别照片和复杂图表中的细微细节。在一个演示中,它在一张巨大的航拍图像中找到了仓库屋顶上一个约两米大小的部件。该模型还能首次操作计算机,自主控制软件。腾讯专门用 Blender 和 CAD 程序等专业工具,以及需要按顺序使用多个应用程序的任务对它进行了训练。
Hy4 通过代码生成裁剪片段,在一张 28 亿像素的航拍图像中搜索屋顶上一个约两米大小的部件。| 图片:腾讯
图像理解还旨在辅助编程,例如构建网站和游戏。在游戏开发方面,腾讯与自家游戏工作室合作,并针对 Unreal、Unity 和 Godot 等主流游戏引擎对模型进行了适配。据开发团队称,在手游《王者荣耀》中,Hy4 几乎能完美处理配置任务,发现漏洞根本原因的频率也远高于以往。
金融成为关键试炼场
腾讯在金融和办公应用上投入了特别的精力。据腾讯介绍,在一项由 200 多名金融专业人士参与的测试中,他们对 Hy4 在近 600 项真实任务上的表现进行了评估,该模型排名领先于中国竞争对手 Deepseek V4.1 Flash、Kimi K3.1 Pro 和 GLM-5.3 Flash。Hy4 在进行网络调研时产生的幻觉也理应大幅减少。得益于改版后的搜索功能,内部测试中的幻觉率从近 16% 降至约 5%。该模型还能在 Word、Excel 和 PowerPoint 文件上与人类实时协作。
合作公司也报告了类似进展,不过它们同样依赖各自的内部测试。金融平台 Comein 认为,Hy4 是唯一能与 Anthropic 的 Claude Opus 5.5 并列第二的中国模型,并在数据查询方面明显领先于 OpenAI 的 GPT-6.1 Sol。智能体平台 Manus 报告称,Hy4 几乎能成功完成所有文档任务。
在腾讯的评估中,Claude Opus 5.5 在全部十二项基准测试中都领先于 Hy4。GPT 6.1 Sol 在其中八项上胜过 Hy4。| 图片:腾讯
Artificial Analysis 的初步独立评估则呈现出略有不同的图景。在其 Intelligence Index 上——该指数综合了知识工作、编程、科学和通用知识——Hy4 得分为 35 分,高于 30 分的中位数。这一成绩与 Grok 4.7 持平,在中国实验室的模型中位列第二,仅次于得分 46 分的 Kimi K3.1 Pro。在腾讯自己的金融测试中,Hy4 仍排在 Kimi 之前。
不过,在不联网搜索的情况下,情况就变了。在 AA-Omniscience 事实准确性测试中,Hy4 依靠其训练数据能正确回答约一半的问题。而当它不知道答案时,有 89% 的概率会编造一个答案,而不是承认自己存在知识空白。
在 Artificial Analysis 的 Intelligence Index 上,腾讯的模型大幅落后于中国最强模型 Kimi 3.1 Pro。它与 Anthropic 最小的模型 Haiku 5.5 大致持平,而后者在该基准测试中每项任务的成本几乎只有一半。| 图片:Artificial Analysis
帮助构建自身下一代版本的 AI
据腾讯称,在一项涵盖七个领域研究任务的内部测试中,Hy4 在开放可用的模型中排名第一。Hy4 还与智能体系统 Hyra 2.0 配合,优化了腾讯自身用于更小模型的训练流程。
据腾讯称,Hy4 搭配 Hyra 2.0 框架后,将公司内部训练基础设施的吞吐提升了约 20%,涉及六个步骤。| 图片:腾讯
随后,它独立训练出了 Jev 的自有版本——Jev 是美国初创公司 TypeSafe AI 的决策模型。该版本在多项未具名的测试中优于原版。腾讯将以开放模型的形式发布两个通过这种方式创建的生物学和物理学专用模型。
由 Hy4 训练的 Hyra-Jev 变体在八项基准测试中的六项上击败了基础 Jev 模型。| 图片:Tencent
展望未来,Hy4 旨在支持并加速公司内部进一步的人工智能研发。
安全性、定价与公开承认的短板
考虑到计划中的权重发布,Tencent 强调安全性。公司表示,其在训练过程中有意过滤掉了与黑客攻击相关的材料,Tencent 自家的安全实验室也未发现滥用风险上升。该模型在生物危害方面也远低于 Tencent 的风险阈值。详细的安全报告将随模型权重一同发布。
尽管有所改进,Tencent 仍按预览版的价格收费。100 万输入 Token 约 0.83 美元,100 万输出 Token 约 2.50 美元。不过公司承认,在高等数学、较长的编程项目、复杂图像分析等困难任务上,Hy4 仍远远落后于最好的商用模型。接下来还计划推出更快、更便宜的变体 Hy4 flash。
差距在 Hy-FrontierMath T5 上最大,Hy4 得分为 13.0 分,而 GPT 6.1 Sol 为 97.5 分,Claude Opus 5.5 为 91.3 分。| 图片:Tencent
Tencent 去年 7 月刚以 Apache 2.0 许可发布了拥有 2950 亿参数的 Hy3。自那以后,中国的竞争对手们纷纷推进。Zhipu AI 发布了 GLM-5.3,Deepseek 发布了 V4.1-Flash-1017,Moonshot AI 推出了 Kimi K3.1 Pro。Artificial Analysis 在 10 月下旬改版了其指数,如今其中 60% 基于私有测试集。开放模型与顶尖模型之间的差距已从 3 分扩大到 12 分。
Hy4 计划发布权重之际,正值规则收紧之时。中国商务部和科技部于10月发布了一份出口管制框架草案,将涵盖任何位列全球前十或参数超过两万亿的模型的模型权重。不过,对于开源发布,该草案包含一个自动批准的快速通道流程。
没有炒作的 AI 新闻——由人类精选
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models)
Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models
Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models
GPT-Image-2 prompted by THE DECODER
Tencent has released its AI model Hy4, which can analyze images and operate software on its own. The model weights will follow at a later date.
The model was specifically optimized for coding, finance, and office work. Tencent says it has already used Hy4 to improve training pipelines for other models.
In independent tests, Hy4 ranks second among Chinese models but falls well behind top models like Claude Opus 5.5 on hard math and coding tasks.
Tencent has unveiled its AI model Hy4. It can understand images for the first time and operate software on its own. The company counts it among the best openly available models, though the model weights won't be released until later.
About two months after the preview, Tencent has released the finished language model Hy4. According to Tencent, the model was trained on significantly more data than its predecessor and "for the first time" ranks among the strongest open-weight models. Until the weights are released at a still-unspecified date, Hy4 is only available through Tencent Cloud, OpenRouter, and Tencent apps like the Yuanbao chatbot.
The biggest new feature is image understanding. Hy4 is supposed to recognize fine details in photos and complex diagrams. In a demo, it finds a roughly two-meter component on a warehouse roof inside a massive aerial image. The model can also operate a computer for the first time, controlling software on its own. Tencent trained it specifically on professional tools like Blender and CAD programs, as well as tasks that require using multiple applications in sequence.
Hy4 searches a 2.8-gigapixel aerial image for a roughly two-meter component on a rooftop by generating cropped sections via code. | Image: Tencent
Image understanding is also meant to help with coding, such as building websites and games. For game development, Tencent worked with its own gaming studios and adapted the model for popular game engines like Unreal, Unity, and Godot. On the mobile game Honor of Kings, Hy4 handles configuration tasks nearly flawlessly and finds the root causes of bugs far more often than before, according to the dev team.
Finance as a key proving ground
Tencent put particular effort into finance and office applications. In a test with over 200 finance professionals who evaluated Hy4 on nearly 600 real-world tasks, the model ranked ahead of Chinese rivals Deepseek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash, according to Tencent. Hy4 is also supposed to hallucinate far less during web research. Thanks to a revamped search feature, the hallucination rate dropped from nearly 16 to about 5 percent in internal tests. The model can also collaborate with humans in real time on Word, Excel, and PowerPoint files.
Partner companies report similar progress, though they're also relying on their own internal tests. Financial platform Comein sees Hy4 as the only Chinese model on par with Anthropic's Claude Opus 5.5 in second place, and clearly ahead of OpenAI's GPT-6.1 Sol on data queries. Agent platform Manus reports that Hy4 completes nearly all document tasks successfully.
In Tencent's evaluation, Claude Opus 5.5 leads Hy4 across all twelve benchmarks. GPT 6.1 Sol beats Hy4 in eight of them. | Image: Tencent
An initial independent assessment from Artificial Analysis paints a slightly different picture. On its Intelligence Index, which combines knowledge work, coding, science, and general knowledge, Hy4 scores 35 points, above the median of 30. That ties it with Grok 4.7 and puts it in second place among Chinese lab models, behind Kimi K3.1 Pro at 46 points. In Tencent's own finance test, Hy4 still ranked ahead of Kimi.
Without web search, though, the picture changes. On the AA-Omniscience factual accuracy test, Hy4 answers about half the questions correctly from its training data. When it doesn't know an answer, it makes one up 89 percent of the time instead of admitting the gap.
On the Artificial Analysis Intelligence Index, Tencent's model trails the strongest Chinese model, Kimi 3.1 Pro, by a wide margin. It roughly ties with Anthropic's smallest model, Haiku 5.5, which costs almost half as much per task on the benchmark. | Image: Artificial Analysis
AI that helps build the next version of itself
In an internal test with research tasks across seven fields, Hy4 ranks first among openly available models, according to Tencent. Together with the agent system Hyra 2.0, Hy4 also optimized Tencent's own training pipelines for smaller models.
According to Tencent, Hy4 with the Hyra 2.0 harness boosts throughput of the company's internal training infrastructure by about 20 percent across six steps. | Image: Tencent
It then independently trained its own version of Jev, a decision-making model from US startup TypeSafe AI. That version outperforms the original on several unnamed tests. Tencent is releasing two specialized models for biology and physics that were created this way as open models.
The Hyra-Jev variant trained by Hy4 beats the base Jev model in six out of eight benchmarks. | Image: Tencent
Going forward, Hy4 is meant to support and speed up further AI development inside the company.
Safety, pricing, and admitted weaknesses
With the planned weight release in mind, Tencent is stressing safety. The company says it deliberately filtered out hacking-related material during training, and Tencent's own security labs found no elevated risk of misuse. The model also falls well below Tencent's risk threshold for biological hazards. A detailed safety report will follow alongside the model weights.
Despite the improvements, Tencent is charging the same price as for the preview. One million input tokens costs about $0.83, and one million output tokens runs about $2.50. The company admits, however, that Hy4 still falls well behind the best commercial models on hard tasks like advanced math, longer coding projects, and complex image analysis. A faster, cheaper variant called Hy4 flash is planned next.
The gap is widest on Hy-FrontierMath T5, where Hy4 scores 13.0 points compared to 97.5 for GPT 6.1 Sol and 91.3 for Claude Opus 5.5. | Image: Tencent
Tencent released Hy3 with 295 billion parameters under the Apache 2.0 license just last July. Since then, Chinese rivals have pushed ahead. Zhipu AI shipped GLM-5.3, Deepseek released V4.1-Flash-1017, and Moonshot AI launched Kimi K3.1 Pro. Artificial Analysis revamped its index in late October, with 60 percent of it now based on private test sets. The gap between open models and the top has widened to 12 points from 3.
The planned release of Hy4's weights comes at a time of tighter rules. China's commerce and science ministries published a draft export control framework in October that would cover model weights for any model ranked in the global top 10 or exceeding two trillion parameters. For open-source releases, however, the draft includes a fast-track process with automatic approval.
AI News Without the Hype – Curated by Humans
原文为英文,正文由自动抓取并保留原文。
原文发布于 The Decoder(Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。