Both versions are aligned block by block, in reading order: title, the essentials, then paragraph by paragraph. Where the translation merged or split a paragraph, the matching cell stays empty — we never pair two passages by guesswork.
聊天窗口里的图片，开始有了新的处理方式。微信正在灰度测试“AI修图”，用户在图片预览页点击“AI处理”，就能进入小微对话框，调用3大类11项功能，美颜、滤镜、文字翻译、自动打码和超分都在其中。按钮没有把人带去另一个应用，修图就发生在原本的聊天路径里。
Images in chat windows are starting to offer a new way to be edited. WeChat is conducting a limited rollout of “AI photo editing.” Users can tap “AI processing” on the image preview page to open a Xiaowei dialog and access 3 categories and 11 features, including beautification, filters, text translation, automatic redaction and super-resolution. The button does not take users to another app; the editing happens within the chat flow they were already using.
爱范儿的测试显示，小微在美肤、扩图、调整比例和消除杂物等编辑任务上表现尚可。瑕疵仍然看得出来，但对于边聊天边改一张图的轻量需求，已经够用。它更像一个随手可用的工具，而不是要替代专业修图软件的完整工作台。
爱范儿’s testing shows that Xiaowei performs reasonably well on editing tasks such as skin smoothing, image expansion, aspect-ratio adjustment and object removal. Imperfections remain visible, but it is already sufficient for lightweight needs such as editing an image while chatting. It is more like a handy tool than a full-fledged workstation intended to replace professional photo-editing software.
差距出现在重绘和生成新内容时。用户需要更清楚地描述想要什么，提示词质量会直接影响结果。通过直接询问小微，爱范儿了解到它目前使用的是微信定制的WeLM模型，大概率是此前已在对话和小程序中上线的WeLM-80B版本，这一判断并非微信正式公布的型号确认。
The gap becomes apparent when repainting or generating new content. Users need to describe more clearly what they want, and prompt quality directly affects the results. By directly asking Xiaowei, 爱范儿 learned that it currently uses a WeChat-customized WeLM model, most likely the WeLM-80B version previously launched in chats and mini programs. This assessment is not confirmation of the model version officially announced by WeChat.
小微的设计重点，也许不在于成为一个醒目的独立入口。微信正尝试把小微拆进既有功能，而不是先打造一个独立的AI应用。腾讯披露，微信团队定制WeLM时重点考虑了微信场景和推理效率；公开的WeLM-HD4-80B和WeLM-HD4-617B总参数量分别达到800亿和6170亿，但实际激活参数只有30亿和230亿。微信及WeChat截至今年6月合并月活达到14.39亿，在这样的用户规模上，每次模型推理都要面对现实的算力账单。腾讯最新季度资本开支达到528亿元，同比增长176%，自由现金流转为负138亿元；腾讯称，这些基础设施正在支持模型、WorkBuddy、CodeBuddy以及微信AI项目的训练和推理。
Xiaowei’s design focus may not be on becoming a prominent standalone entry point. WeChat is trying to integrate Xiaowei into existing features rather than first building an independent AI app. Tencent has disclosed that the WeChat team focused on WeChat-specific scenarios and inference efficiency when customizing WeLM. The publicly disclosed WeLM-HD4-80B and WeLM-HD4-617B have total parameter counts of 80 billion and 617 billion, respectively, but only 3 billion and 23 billion active parameters. As of June this year, the combined monthly active users of WeChat and WeChat reached 1.439 billion. At this scale, every model inference has to contend with a real computing bill. Tencent’s latest quarterly capital expenditure reached 52.8 billion yuan, up 176% year over year, while free cash flow turned negative at 13.8 billion yuan; Tencent said this infrastructure is supporting the training and inference of models, WorkBuddy, CodeBuddy and WeChat AI projects.
具体来说，用户现在能获得的是更顺手的轻量图片编辑，而不是一次完成的高质量生成。微信的价值在于把小微放进人们已经每天使用的图片、搜索和内容场景，降低尝试AI的门槛；代价则是效果需要继续打磨，且成本压力已经反映在财务数据中。腾讯最新季度Non-IFRS营业利润为756亿元，剔除元宝、WorkBuddy、CodeBuddy、小微等AI产品影响后为861亿元，按腾讯自己的口径，新AI产品目前拖累当季经营利润约105亿元。微信这次灰测，正是一个小功能如何成为大规模AI基础设施入口的切面。
Specifically, what users can get now is more convenient lightweight image editing, not high-quality generation completed in a single step. WeChat’s value lies in putting Xiaowei into the image, search and content scenarios people already use every day, lowering the barrier to trying AI. The trade-off is that the results still need refinement, while cost pressures are already showing up in the financial data. Tencent’s latest quarterly Non-IFRS operating profit was 75.6 billion yuan, or 86.1 billion yuan excluding the impact of AI products including Yuanbao, WorkBuddy, CodeBuddy and Xiaowei. By Tencent’s own calculation, the new AI products currently reduced operating profit by approximately 10.5 billion yuan in the quarter. WeChat’s latest limited rollout offers a glimpse of how a small feature can become an entry point to large-scale AI infrastructure.