文丨海克财经 许俊浩
全球AI持续迭代，头部玩家仍在沿着模型与产品两条线并行提速。
国内互联网大厂腾讯近期频密动作已引得业界更多关注。腾讯日前发布并开源的混元同声传译大模型Hy-SimulMT即在此列。据官方资料，Hy-SimulMT在翻译高准确率的同时，做到了低时延，平均响应时间仅为3秒。
混元大模型的底层进阶已颇显著。在AI大模型聚合平台OpenRouter全球调用量榜单上，2026年10月27日当周，腾讯混元大模型Hy4 preview调用量为9万亿tokens，排名第五，位列DeepSeek V4.1 Flash 1017、DeepSeek V4.1 Flash、智谱GLM-5.3 Flash和MiniMax M3.1 Flash之后；而在2025年，混元系列模型甚至无法排入OpenRouter调用量榜单。
作为目前全球最大的大模型聚合平台，OpenRouter所列榜单尽管统计的只是通过该平台的大模型调用量，但已具备相当的说服力。国内一些出色的模型凭借更高的性价比和在此基础上更大的使用量，在榜单上或可超过部分海外头部模型。虽然使用量不等同于模型智能水准，但可由此窥得一些端倪，比如时下混元大模型已肉眼可见地在稳中求进，并逐步拉高用户心智。
腾讯在AI产品方向也在加快追赶的脚步。除了继续推进元宝、WorkBuddy等已初获声量的产品外，腾讯对更多新意的倾力下注也已被陆续感知，其中包括已被曝出的对个人智能体（Personal Agent）产品Handy Bot的开发。
2026年6月开始灰度测试的微信原生AI助手小微尤需重点提及。国民级应用微信天然具备打造爆款乃至孵化超级入口的能力，因为一款产品或一个功能一旦在微信生态内跑通和崛起，也就意味着它经受住了近乎全民考验而朝着通用型行业基建而去，潜力自不必说。
调研机构QuestMobile数据显示，2026年6月，微信APP MAU（月活跃用户数）为11.54亿，位居国内APP用户规模第一，抖音APP以10.29亿排在第二；2026年9月，微信APP MAU进一步增长至11.66亿。
获得小微内测资格的用户，微信消息列表左上角往往会出现一个眼睛式图标，点进去是标注测试版的小微对话界面。目前小微仍处测试阶段，未予全量推送，但功能在持续开拓，可唤起的位置也已从首页，扩展到了公众号、朋友圈、聊天对话框等多个触点，能够发消息、调设置、点外卖、调用小程序。
微信灰测小微让很多人深感似曾相识。8年前的短视频一战，腾讯旗下微视APP等一众平庸产品，面对字节旗下抖音的狂飙突进,节节败退、无力招架，最终是微信临危受命、接过重任，于2020年1月灰测视频号并力推后者稳健爆发，为腾讯在该赛道逆势拿下了行业前三的顶流卡位，挽回了些许颜面。
小微在称谓上虽以“小”为前缀，但它所承载的使命，则绝不是一个“小”字所能遮蔽的——它指向的是以微信生态为基本盘的AI超级入口。
对照视频号演进逻辑，接下去透过小微入口，目前市面可见的所有聊天机器人(Chatbot)或个人智能体等产品能够实现的功能，都或有可能在这里全部寻得新解，腾讯AI或将由此迎来新的可能，行业格局或将因之一变。
01 藏身微信生态
小微本质上是AI与社交的结合。
如果借用此前“互联网+”的说法，那么当下局面更像是“AI+”，每个具体场景都在被重新审视，看AI如何介入与结合。在AI领域，智能体又比聊天机器人更进一步，不仅能提供问答，还能拆解任务、调用外部应用，代用户完成发消息、订服务、购物等操作。
微信的标签是社交，置身微信生态的小微则起于微信社交，它先连接聊天、朋友圈、公众号等高频场景，再向小程序服务延伸，大体路径是让智能体渐渐全面触达微信所能触达的用户全部日常生活场景。
目前小微已阶段性融入微信大部分功能。除首页左上角外，在聊天界面下拉菜单、公众号文章详情、文章转发界面等都已出现“问小微”入口，“扫一扫”中也加入了识别、翻译、记住此刻等功能。
在灰测迄今的4个月里，小微的授权边界在逐步放宽。据海克财经观察，原先用户如果想要使用小微代为发送消息或朋友圈，小微会回复只能帮忙编辑文案，无法代为发送，现在则能够准备好所有前置信息，由用户跳转相关页面确认后发送；用户也可借小微向好友定时发送消息，只是消息发出后，双方聊天界面中该条信息下方均会标注“通过小微定时发送”。
与微信深度绑定的小微，能够拿到微信社交数据权限，包括读取聊天记录、文档文案、朋友圈内容等，这是国内其他智能体难以比拟的生态长板。用户尚未发问，小微便会主动提醒你的朋友在看什么文章。它也能帮用户分析聊天记录，且为保护隐私，小微会提示聊天记录仍保存在本地，自己只调用当天内容。
小微未来高光有望由此而来。智能体的智能表现，既取决于底层大模型的智能水平，也受上下文（Context）影响。没有上下文的智能体，就等于一张白纸，需要用户额外提供大量信息。微信握有足够的上下文，包括聊天记录、朋友圈、公众号阅读、小程序使用轨迹等，这些构成了一套持续、真实、高频的用户画像。基于该画像，小微对用户的了解，不必依赖额外调教，也无需用户反复交代背景，智能体可以直接在既有语境里理解需求、判断意图。
微信庞大而丰富的小程序生态亦需重点提及。微信多年来通过小程序已将海量服务接入阵营，这些服务近乎涵盖移动互联网所有类别，用户日常所需购物、出行、挂号、缴费、订票等均能调用这些小程序完成。小微灰测上线前，腾讯为微信接入智能体铺平道路，发布了《关于开发者接入微信AI生态的指引》，美团、携程、同程、滴滴等作为首批内测应用，与微信团队联合开发并测试接入。
目前小微能够掌控的是前置信息设置的环节，后续还有一些步骤仍要用户主动确认、手动完成。比如小微代发红包，给出对象、金额与祝福语后，小微需要用户先确认联系人是否正确，再跳转到红包页面，让用户手动发送。但这些实属产品打磨的必经阶段，入口、服务、授权链路一旦打通，小微能调动的就不只是聊天框里的几项功能，而是整个微信生态的服务网络。
02 扛起腾讯重压
小微动作可上探至腾讯战略。
自2018年“930变革”迄今，腾讯在组织层面一直保持着六大事业群格局，目前模型供给以技术工程事业群（TEG）为主，元宝、WorkBuddy等产品归于云与智慧产业事业群（CSIG），小微则归于微信事业群（WXG）。
WXG自始至终由腾讯副总裁、微信创始人张小龙主导，而张小龙主导的WXG在腾讯体系内是个众所周知的鲜有的相对独立的存在。而在AI方向，微信目前立足自研，与混元大模型无涉。据官方资料，小微底层模型是微信团队自研的WeLM。小微答问也提到，它的底层主模型为WeLM，部分回答由DeepSeek承接。
事实上，WeLM研发颇早，微信团队早在2022年10月既已发布相关信息，ChatGPT引爆全球是一个月之后的事。小微内测后，WeLM随之进入大众视野，市场多将其解读为面向微信场景训练的低成本小模型。2026年7月，WeLM官方技术团队还开设小红书账号，对外披露技术进展。
尽管外界质疑小微不接入混元大模型是重复造轮子，但这既涉及超大体量的微信对容错率的极低接受度，又关乎WXG团队技术实现能力及决策自主性，还与腾讯内部颇有荣光的独特赛马机制强相关。举例来说，2010年前后，移动IM（即时通讯）酝酿早期，手机QQ团队、张小龙QQ邮箱团队等腾讯内部三四个团队同时开启相关APP研发，最终微信胜出。
前述短视频一役有着相似背景。腾讯旗下短视频APP微视起步并不晚，微视2013年9月上线时，快手上线不过2年，抖音则要迟至3年后才上线，但微视此后不温不火，甚至一度宣布要关停。抖音上线后的爆发式增长让腾讯备感焦灼，后腾讯于2018年重启了微视，同时推出了下饭视频、速看、哈皮等多款相似短视频APP。奈何这些跟随型产品效力不彰，腾讯最终通过微信在2020年推出视频号的方式扳回一局。
腾讯AI同样承压日久。AI大模型自2022年11月ChatGPT上线后迅速掀起全球浪潮，国内玩家纷纷推出自家大模型和相应C端产品，百度于2023年3月率先基于文心大模型发布文心一言APP，1个月后阿里发布基于千问大模型的通义千问APP，同年8月字节发布基于豆包大模型的豆包APP，腾讯则在2023年9月发布混元大模型，2024年5月才推出基于混元大模型的元宝APP。彼时元宝产品表现平平，混元大模型在技术上也未展现出强劲竞争力，最初甚至无力进入SuperCLUE、C-Eval、OpenCompass等大模型主流评测榜单。
面对尴尬窘境的腾讯开始了低调逆势发力。前OpenAI研究员姚顺雨加入腾讯任首席AI科学家后，腾讯在2026年3月撤掉成立了近10年的AI Lab，将混元的模型、算力、数据部门逐步收拢，交由姚顺雨统一负责，归于TEG。调整之后，腾讯于2026年4月发布并开源新混元大模型Hy3 preview，该模型连续3周登顶OpenRouter调用量榜单，直到5月被DeepSeek V4 Flash超越。
但即便如此，目前腾讯AI模型与产品的业内站位，尚难匹配腾讯业界巨头能量，特别是模型智能水准依旧未能进入全球第一梯队。据大模型评测平台Artificial Analysis 2026年10月27日实时智能指数榜单，Hy3不在前30名之内，与海外友商Anthropic旗下Claude系列、OpenAI旗下GPT系列差距悬殊，也落后于国内Kimi K3.1、阿里Qwen3.8、DeepSeek V4.1 Flash等。腾讯压力可以想见。
而在产品方向，腾讯亟需一个AI超级入口，但四顾茫然。曾在2026年2月春节期间斥重金补贴引流的元宝，如今存在感已大幅回落；而在AI办公领域表现突出的WorkBuddy，目前用户体量相较更具想象力的国内C端用户大盘仍微不足道，而且赛道不同，WorkBuddy对腾讯争夺AI超级入口作用有限。
QuestMobile数据显示，依靠腾讯体系支援和补贴的元宝，2026年春节MAU冲到超过1亿，2026年9月已仅剩3733万，而当月排在国内AI原生APP前三名的分别是豆包（4.1亿）、千问（1.5亿）、DeepSeek（1.2亿），同月在国内AI办公领域高居榜首的WorkBuddy，MAU为3235万。
小微出场因此被寄予厚望。
03 或可突出重围
腾讯AI产品已多线铺开。
据海克财经了解，目前腾讯AI已上线的产品，在元宝、WorkBuddy、小微之外，还有PCG应用宝团队推出的个人AI助手马维斯（Marvis），CSIG旗下QQ浏览器团队推出的AI知识管家ima等；CSIG推出的Buddy系列甚至已形成一个家族，除了最早的CodeBuddy和从CodeBuddy衍生出的WorkBuddy，另有LearnBuddy、MusicBuddy、DataBuddy等不同方向的产品。
产品多并不必然导向成功。腾讯AI产品能真正走入大众视野的是极少数，比如与WorkBuddy在2026年3月同时发布的QClaw。据财新等多家媒体报道，在2026年1月龙虾即OpenClaw点燃全球智能体热情后，腾讯内部各团队可申请算力，基于龙虾开源生态做自己的龙虾产品，WorkBuddy、Qclaw均在其中。但仅运营半年，QClaw就在2026年9月发布了停止运营公告。
小微有着丰富的探索可能。以服务为例，目前小微所处阶段是识别用户意图、唤起某个具体小程序；接下去小微或有可能坐镇微信主页面，定位总管型智能体，比如当微信聊天界面收到出差指令，小微便可通过上下文理解行程安排，再调取OTA小程序预订交通票务和酒店住宿，再用打车软件衔接机场到酒店的路程，最后由微信支付完成闭环。
腾讯对小微期待颇高。在2026年8月腾讯财报发布后的电话会议上，腾讯总裁刘炽平提到，腾讯构想的未来是，大量用户通过小微及各种智能体来执行指令并逐步完成交易；过去微信生态是用户自己与内容、与小程序交互，未来用户只需向智能体下达一条复杂指令，智能体就能帮用户执行交易；更长远地看，每个用户都会拥有自己的智能体，智能体之间可以相互通信并完成交易。刘炽平的表述可直白理解为，全民标配智能体的时代渐行渐近，小微已朝担纲该角色迈进。
微信已为之蓄力多时。以支付为例，2026年6月，微信支付正式发布AI专属卡，这是一项专为AI智能体支付场景设计、内置于微信零钱的功能。与授权AI调用主账户的方式不同，该卡用户可为AI智能体在微信零钱中单独开通一张专属卡并自行将资金转入其中。AI专属卡不获取用户账户密码等关键信息，不直接动用主账户资金，而是在用户设定的范围内、经用户许可后完成购买与消费。AI专属卡首发接入的是WorkBuddy，后续接入小微并不复杂。
业界对小微大多抱持乐观态度。长江证券2026年8月31日发布的一份研报提及，微信的聊天对话场景或可重构智能体入口，将微信打造为社交场景的AI智能中枢，借助小程序与自身应用体系，丰富智能体功能、激活生态并推进后续商业化；腾讯AI进展加速，微信小微打通入口到交互再到履约的路径后，微信会由超级APP转向智能体操作系统，模型-产品-生态三层框架有望构建不可替代的服务闭环。国际投行花旗在2026年9月25日发布的研报中，则用“终极消费AI代理”来指称小微。
腾讯AI压力已然给到微信，微信已力推小微踏上探索之路。这是小微的一小步，腾讯的一大步。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen Wen Wen Wee Kee, Xu Jun-ho
Global AI continues to overlap, and head players continue to accelerate along both models and products.
Tencent, a large domestic Internet plant, has recently attracted more attention from the industry. This is the case of LLMHy-SimulMT, a hybrid with open sources. According to official sources, Hy-SimulMT, while translating with high accuracy, achieved a low latency response time of only three seconds on average.
The bottom steps of the hybrid LLM have been quite significant. On the AILLM polymer platform OpenRouter ' s global call list, on the week of 27 October 2026, the Tencent hybrid LLMHy4 preview was 9 trillion tokens, ranked fifth, in DeepSeek V4.1 Flash 1017, DeepSeek V4.1 Flash, ZhipuGLM-5.3 Flash and Mini Max M3.1 Flash; in 2025, the hybrid series model was not even able to fit into OpenRouter ' s call list.
As the largest LLM platform in the world at present, OpenRouter's list, although measured only through LLM of the platform, is quite convincing. Some of the country ' s excellent models, based on higher value for money and greater use on that basis, may be on the list to exceed some of the overseas head models. Although the amount of use is not equivalent to the model ' s intelligence level, it provides some insight into, for example, the fact that the time mass LLM is now visiblely moving forward in a steady and gradual way and is gradually raising the user ' s mind.
Tencent is also accelerating catch-up in the direction of AI products. In addition to continuing to advance products such as Yuanbao, WorldBuddy and others, Tencent's more innovative dumping has been detected, including the exposed development of the individual AI agent (Personal Agent) product, Handy Bot.
The WeChat Native AI assistant, who began the greyscale test in June 2026, needs special mention. The application of WeChat to the national level naturally has the capacity to create a burst or even hatch a super-entry, because once a product or a function runs and rises within the WeChat ecology, it means that it has survived the near-universal test and is moving towards the building blocks of a generic industry.
The research agency QuestMobile shows that in June 2026 WeChatAPP MAU (number of active monthly users) was 1,154 million, ranked first in the country in terms of APP users and second in terms of DouyinAP, with 1,029 million; in September 2026, WeChatAPP MAU increased further to 1,166 million.
Users who are eligible for small intrometers, the upper left corner of the WeChat message list tends to have an eye-like icon, with the point of entry being the micro-dialogue interface that points to the test version. It is currently in the testing phase and is not being delivered in full, but it is being developed continuously, and its callable location has been extended from the front page to multiple contact points, such as WeChat Official Account, the circle of friends, the chat dialog, to send messages, set up, order take-out, and call a small application.
WeChat ash is a small measure that many people seem to know. Eight years ago in the short video World War I, under the Tencent flag, a crowd of mediocre products, such as the APP, went up in the face of byte ' s Z Q58 Z raging, stifling down, unable to fight, and finally, WeChat was in danger and overburdened, in January 2020, as a result of ash-detecting and pushing the latter into a steady explosion, thus taking Tencent ' s top three positions in the industry ' s opposition on that track, thus restoring some face.
Even though it is prefaced by the word "small," the mission it carries is by no means hidden by the word "small" - It points to the WeChat eco-based AI Superport.
By using video numbers to evolve logic, then through the mini-ports, all the functions that can be achieved by all the chat robots (Chatbot) or individual AI agent products that are currently visible on the market are, or are likely to be, all of them found here. Tencent AI may be able to open up new possibilities and the industry pattern may change.
WeChat Ecology
It's actually an AI-social combination.
If the term “Internet+” is used, then the situation is more like “AI+”, and each specific scene is being re-examined to see how AI intervenes and combines. In the area of AI, AI agent goes further than chat robots, not only to provide questions and answers, but also to dismantle missions, call external applications and perform messaging, booking services, shopping, etc. on behalf of users.
WeChat labels are social, and WeChat micro-cosms start with WeChat, which connects high-frequency (HF) scenes such as chats, friends ' circles, WeChat Official Account, then expands to small programs, the broad path being to enable AI agent to reach all users of WeChat's daily life.
At present, WeChat functions have been incorporated in a small and phased manner. In addition to the top left corner of the front page, the entry "Quest" has appeared under the chat interface, details of WeChat Official Account articles, articles forwarding interfaces, etc. The "Sweep" has been added to the "Scanning".
In the four months since the grey survey, the authorized border has gradually been eased. According to Hicks, if the original user wanted to use the mini-decease to send a message or a circle of friends, it could only help to edit the file and could not send it on its behalf, and was now able to prepare all the pre-messages that the user could send by jumping to the relevant page for confirmation; the user could also use the small-micro-message to send a message to a friend on a regular basis, except that, after the message had been sent, the message would be marked “through a small-timed time”.
The WeChat deep-seated microcosm that allows access to WeChat social data permission, including reading chat records, document files, friends ' circles, etc., is an ecological long board that other AI agents in the country cannot compare. Users have not yet asked what their friends are reading. It also allows users to analyse chat records and, in order to protect their privacy, they are kept locally and they call only the same day.
It is hoped that the light of the small future will come. The intelligence performance of AI agent depends on both the intelligence level of the bottom LLM and the influence of context (Context). Without context, AI agent is equal to a piece of paper, requiring a large amount of additional information from users. WeChat has enough context, including chat records, friends ' circles, WeChat Official Account reading, small program trajectories, etc., which constitute a continuous, real, high-frequency user image. Based on the image, the understanding of the user is not dependent on additional instruction, nor on the user ' s repeated background, and AI agent can understand the need and the intent directly in the established context.
WeChat’s large and rich small program ecology also needs to be highlighted. WeChat has, over the years, been able to access large-volume service access camps through micro-programs, which almost cover all types of mobile Internet, and which users can use for their daily shopping, travel, registration, fees, booking, etc. Tencent paves the way for WeChat to access AI agent, publish to the WeChat AI Ecology Guidelines for Developers to Access WeChat AI, and WeChat to develop and test access in conjunction with the WeChat team as the first internal measurement applications.
What is currently under control is the front-line information setup, and a number of subsequent steps remain to be confirmed and completed manually by the user. For example, after giving objects, amounts and blessing, the user is required to confirm the contact ' s correctness and then jump to the red bag page to allow the user to send it manually. These are, however, essential stages in the grinding of products, and once access, services and authorized links have been reached, micro-mobilization is not just a few functions in chat boxes, but an entire WeChat eco-service network.
02, carry Tencent.
A little action can be traced to Tencent strategy.
Since the “930 Change” in 2018, Tencent has maintained a six-constituency pattern at the organizational level, with the current model supply being dominated by the TEG, and products such as Yuanbao, WorkBuddy and others attributed to the Cloud and Wisdom Industry Enterprise (CSIG) and the WeChat Enterprise (WXG).
Zhang Xiaolong, Vice President of Tencent and founder of WeChat, was the leader of WXG, a well-known relatively independent presence within the Tencent system. In the direction of AI, WeChat is now self-studyed and has nothing to do with mixed LLM. According to official sources, the microdemography is the WeLM self-study by the WeChat team. It was also mentioned that its bottom main model is WeLM, partially answered by DeepSeek.
In fact, WeLM was developed early, the WeChat team had already had information on publish in October 2022, and ChatGPT was a month later. After micro-intrometry, WeLM entered the public view, and the market interpreted it as a small, low-cost model for WeChat field training. In July 2026, the official WeLM technical team also opened an account number REDnote for external disclosure of technological progress.
While the outside world has questioned the fact that LLM, a mixed LLM, is a duplicate wheel, this involves both the extremely low acceptance of the oversized WeChat for error, and the WXG team ' s technical ability and decision-making autonomy, as well as a strong link to the unique and prestigious horse racing mechanism within Tencent. For example, around 2010 the mobile IM (immediate communication) developed early in life, with three or four teams inside Tencent, including the QQQQQ team and the Z Xiaolong QQQbox team, simultaneously initiating the development of the relevant AP and eventually winning WeChat.
The aforementioned short video campaign had a similar background. It is not too late for Tencent under the flag of short videoAPP to begin with, but only two years before Douyin becomes online when September 2013 comes online, but it is not too hot since then, even once declared closed. After Douyin’s explosive growth on the line has simmered Tencent, after which Tencent restarted its microvision in 2018, while many items like short videoAP were launched, such as a food video, a quick look, and Hapi. These follow-on products were ineffective, and Tencent finally pulled back a game by WeChat in 2020.
Tencent AI is also under pressure. AZLLM has rapidly launched a global wave since ChatGPT came on line in November 2022, with domestic players launching their own LLM and corresponding C-end products, Baidu pioneered ErnieLLMZpublishZ7ZAP in March 2023, publish based on LLM, QwenAP one month later, Z113ZZZpublish based on DoubaoAP, Tencent based on DoubaoAP, based on Doubao43Z based on Doubao43Z, Tencent based in September 2023 and LLM based on mixed LLM until May 2024. PEE ' s product was flat, and the mixed LLM did not show strong technical competitiveness and was initially unable to even enter the mainstream LLM ratings of SuperCLUE, C-Eval, Opencompass and others.
Tencent has begun to play a low-key role in the face of embarrassment. Having joined Tencent as chief AI scientist, a former Open AI researcher, Tencent withdrew the almost 10-year-old AI Lab in March 2026, bringing together the hybrid model, calculus, data sector under the unified responsibility of Yao Sun Rain to TEG. Following this adjustment, in April 2026, Tencent and the open-source new hybrid LLMHy3 preview, the model came to the top of the OpenRouter call list for three consecutive weeks until May when it was exceeded by DeepSeek V4 Flash.
Even so, it is difficult to match the current industry position of Tencent AI models and products with the Tencent industry ' s power, especially the model ' s intelligence level, which still fails to reach the first global ladder. According to the LLM evaluation platform Artificial Analysis, the real-time intelligence index list of 27 October 2026, Hy3 is not among the top 30, and the gap between the Claude series, the Open AI GPT series, and the company Anthropic overseas lags behind Kimi K3.1, Ali Qwen 3.8, DeepSeek V4.1 Flash and others. Tencent pressure can be seen.
In the direction of the product, Tencent desperately needed an AI super-entry, but it was not clear. There is now a significant retrenchment of the beauties of the heavy subsidy during the spring of February 2026; and the worldBuddy, which is a prominent user in the AI office, still has a small, more imaginative, domestic C-end user disk, and has a different track, which has a limited effect on Tencent for the AV.
According to QuestMobile data, the dollar treasures that rely on the Tencent system for support and subsidies, in spring 2026, MAU ran to more than 100 million, leaving only 3.73 million in September 2026, while the top three members of the country ' s AI APP were Doubao (410 million), thousands of questions (150 million), DeepSeek (120 million), and WorldBudy, which was the top member in the country ' s AI office for the same month, MAU was 32.35 million.
As a result, the presence of microentrepreneurs was given high hopes.
03 Or highlight the revolving
Tencent AI products have been spread over multiple lines.
According to Hecquest, Tencent AI products are now on line, in addition to Yuanbao, WorkBuddy and Xiao Xiao, as are the personal assistant to the PCG application team Marvis, and the AI knowledge housekeeperma from the CSIG team of the QQQ Browser; the Buddy series launched by CSIG have even formed a family, in addition to the first CodeBuddy and WorldBuddy, derived from CodeBuddy, and products in different directions, such as LearnBuddy, MusicBuddy, DataBuddy.
Many products do not necessarily lead to success. Tencent AI products are a very small number of people who can actually get into the public domain, such as QClaw with WorldBuddy in March 2026. According to various media reports, in January 2026, when the lobster, OpenClaw, ignited the global enthusiasm of AI agent, Tencent internal teams could apply for the calculus to make their own lobster products based on the Lobster Open Source Ecology, in which WorkBuddy and Qclaw were all involved. However, for only six months, QClaw issued a cessation of operations announcement in September 2026.
There are plenty of possibilities for exploration. In the case of services, the current phase of micro-microsets is to identify the intent of the user and to call up a specific micro-program; then go on or possibly sit on the town WeChat main page, and locate the main tube AI agent, for example, when WeChat chat interface receives a travel order, so that the micro-micro can understand the itinerary by context, then call the OTA small program to book traffic tickets and hotel accommodation, then connect the airport to the hotel with a cabling software, and finally close the loop with WeChat payment.
Tencent has a high expectation of smallness. At a telephone conference in August 2026 after the Tencent financial paper publish, President Liu Xinping of Tencent mentioned that the future envisaged by Tencent was for a large number of users to implement directives and gradually complete transactions through micro and various AI agents; that the Z60Z ecology in the past had been for the users themselves to interact with content and with small programs, and that future users would need only to give a complex directive to AI agent, which would allow them to carry out transactions; and that, in the longer term, each user would have its own AI agent, AI agent42Z, which could communicate with each other and complete transactions. Liu Xinping's presentation can be understood as the era of popular AI agent is approaching, and Xiao has moved towards that role in the Platform.
WeChat has been working for a long time. In June 2026, in the case of payments, WeChat paid the official publish AI identity card, a feature designed to pay for AZAI agent, which is embedded in WeChat change. Unlike the way AI is authorized to call the main account, the card users can open a separate card for AZAI agent in WeChat and transfer the funds to them. The AI-exclusive card does not access critical information such as user account passwords and does not directly use funds from the main account, but rather completes the purchase and consumption within the limits set by the user and with the user ' s permission. The first access to the AI-specific card is from WorldBuddy, and the following access is not complicated.
Industry is optimistic about the small majority. A study of publish, dated 31 August 2026, referred to WeChat's chat and dialogue scene or re-engineered AI agent portal, which transformed WeChat into an AI smart hub for social scenes, enriching AI agent's functionality, activating ecology and promoting subsequent commercialization with small programs and its own application system; and Tencent AI's progress accelerated, with WeChat's small access to the interactive Z42Z operating system, where Z60Z would move from superAP to AI agent's operational system, with the model-product-ecological three-tier framework expected to build an irreplaceable service closed ring. In its study of publish, dated 25 September 2026, Citibank International used the term “end-of-consumption AI agent” to refer to smallness.
The pressure of Tencent AI has given WeChat, and WeChat has taken a small step to explore. It's a small step, a big step of Tencent.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.