文 | 镜相工作室，作者丨黄依婷，编辑丨胡苗
一款“龙虾”准备退场，下一轮个人助理（Personal Agent）竞争已经开始。
9月24日，腾讯QClaw宣布，因业务发展调整，将于2026年12月24日停止运营，并提供工具，帮助用户将数据迁移至腾讯另一款产品WorkBuddy。QClaw从3月20日公测到宣布停服安排，只过去了188天。
另一边，大厂正在紧锣密鼓地布局新赛道。10月13日，阿里千问宣布即将发布个人助理产品“千问Go！”，尝试帮助用户处理航班延误理赔、购物比价、电话预约等事务。此前，字节跳动也被曝正在推进个人助理产品，相关项目曾用代号“Spell”。
这样的场景在过去四年间反复出现。我们梳理了腾讯、阿里巴巴、百度、字节跳动，以及月之暗面、智谱、MiniMax、DeepSeek八家公司，在十条主要AI赛道上的产品动作。
数据呈现出两个核心特征：一是竞争周期极速缩短，从文生图、ChatBot到视频生成、AI Agent，新风口接连不断；二是产品生命周期变短，改名、整合、停运成为常态，独立产品正不断向通用助手和底层平台收拢。
公司不断拓宽AI能做的事情，也不断矫正哪些事情值得单独做成一款产品。
四年十赛道：风口越来越多，大厂跟进速度越来越快
过去四年，标志性AI产品的走红，往往会带动整个行业重新分配注意力，这些变化在时间轴上留下了清晰的印记。
2022年，Midjourney、Stable Diffusion等文生图工具受到关注，“输入一句话，生成一张图片”成为许多人认识生成式AI的起点。国内企业迅速跟进，百度的文心一格、阿里的通义万相等产品相继落地，AI正式进入绘画、设计和内容创作等场景。
随后，大语言模型ChatGPT将这股热潮推向更广泛的人群。2023年，百度的文心一言、阿里的通义千问、字节的豆包、月之暗面的Kimi等产品密集亮相。聊天框逐渐成为AI应用的基础形态，搜索、办公、陪伴等不同需求，也围绕它展开。
到了2024年，行业同时向两个方向探索：一边是多模态视频生成，字节的即梦、快手的可灵、智谱的清影、MiniMax的海螺视频等产品接连出现；另一边是AI硬件，阿里、百度、字节等公司试图将智能助手装进耳机、AI眼镜等设备，争取手机App之外的用户入口。
AI热门赛道一浪接一浪，竞争周期越来越短，各家公司的跟进速度也越来越快。
以“龙虾”为例，2026年1月至3月，各种“龙虾”接踵而至。仅3月，便有字节、腾讯、智谱、百度、阿里五家公司，集中公开了自家的“Claw类”产品。对比早期从ChatGPT发布到国内大模型跟进所经历的漫长数月，如今新热点留给公司的观望时间已所剩无几。
这些浪潮并非简单的“依次替代”。新方向不断出现的同时，旧方向的投入仍在继续。公司面临的已不仅是寻找下一个风口，而是要在不断叠加的布局中决定：哪些产品值得继续投入，哪些能力可以合并，哪些试验应该结束。
大厂“全都要”，创业公司“押重点”
在此次统计的十条赛道中，阿里巴巴、字节跳动覆盖全部十条，腾讯、百度各覆盖九条。相比之下四家新秀AI公司中，月之暗面覆盖六条，智谱五条，MiniMax四条，DeepSeek仅两条。
这种差异的背后，是公司资源禀赋和业务重心的不同。对互联网大厂而言，AI不仅是开辟新业务，更是改造原有生态的工具。无论搜索、内容、办公、电商还是硬件，庞大的既有用户和业务场景为它们提供了充足的试验田，也让它们有更多需要守住的市场。
而根基尚浅的AI公司更需要在战略上做出取舍：决定把核心资源优先投向哪里。
赛道铺开之后，哪些产品还留在场上？
进入一条赛道，并不意味着能持续保有一款独立产品。从统计图中可以看出，在赛道铺开后，各家都在不断调整产品的形态与去留。
基础入口全员保留，但其他赛道频繁洗牌。
八家公司全部布局了聊天助手，这是承接搜索、创作和任务处理的基础入口。但在其他方向，产品退出、停运时有发生。
以AI陪伴为例，过去四年，MiniMax的Glow下架，月之暗面的Ohai停更，腾讯音乐的未伴、百度的月匣停运。但同一赛道内，Talkie、星野、猫箱等产品仍然正常运营。
大厂在广撒网后，普遍走向内部整合与收拢。
字节的特点是布局广，同时不断调整产品之间的分工。豆包、即梦、猫箱分别承接不同需求，星绘团队并入豆包，MarsCode编程助手改为TRAE插件，TRAE团队后来也并入了豆包体系……
百度部分能力向既有体系收拢。文心一格迁入文心一言；文心快码并入百度搭子；聊天助手经历“文心一言——文小言——文心”的更名与入口迁移。它一边继续参与多个方向，一边减少部分独立入口。
阿里与腾讯同样存在整合现象。阿里的“悟空”于2026年7月整合进千问办公；腾讯的QClaw已公告将于2026年12月停服，并引导用户迁移至WorkBuddy。
创业公司的路径则出现分化。
DeepSeek的消费产品布局相对集中。与其他公司不断推出新名称、新入口相比，它主要围绕智能助手增加能力。智谱的文生视频产品“清影”并入了“智谱清言”；MiniMax的通用助手“海螺AI”后来更名为MiniMax，原品牌转向视频业务。
产品停运或整合，不仅涉及用户需求的变化，市场竞争的加剧，也源于公司内部的资源取舍。发布时的布局，只是一个起点，持续投入之后，公司才能决定哪些产品值得独立存在。
六款典型产品的轨迹：改名、业务剥离与合并
观察六个典型产品的迭代轨迹，一个反复出现的动作是：改名。
百度的文心一言先改为文小言，后来再改为文心；阿里的通义千问先简化为通义，之后又改为千问。从产品向公众开放到进行第二次更名，两者分别历经了1029天和793天。
名字的变化，往往伴随着公司对产品定位和业务边界的重新安排。
Minimax的海螺AI上线约十个月后，其通用助手更名为MiniMax，原有的“海螺AI”品牌则转向视频业务。名字仍然存在，但所指向的主要用途已经改变。
字节的MarsCode编程助手在国内发布约十个月后，更名为TRAE插件。随着TRAE独立开发环境推出，原来的编程助手被放进更大的产品体系，成为其中一种使用方式。
夸克沿着原有产品持续改版。从2024年7月升级“超级搜索框”，到2025年3月发布“AI超级框”，相隔246天。它尝试把用户熟悉的搜索入口，扩展为能够处理更多任务的AI应用。
当然，也有产品并未等来“改名”的机会。腾讯的QClaw在2026年3月开启公测，到9月便宣布了停服安排（计划于12月停止服务），从公测到宣布停服仅过了188天。不到半年的生命周期，折射出新赛道里产品试错的容错率正在急剧降低。
对公司而言，整合可以集中投入、减少重复建设；对用户而言，频繁改名和迁移，也意味着重新识别入口、适应操作方式。
通用助手越来越像个“功能大杂烩”
独立产品不断分化、整合，通用助手里能做的事情则越来越多。
从聊天，到联网搜索、文件理解，再到绘图、语音、深度研究、生成文档和执行代码，原本被分别讨论的产品方向，逐渐装进了同一个助手。
四款样本的公开记录显示，到2024年末，豆包、千问、DeepSeek和Kimi已经共同具备聊天、联网搜索、文件理解三项能力。此后，研究、办公和执行类能力继续进入部分助手。到2026年10月，豆包、千问、Kimi都已上线文档生成和代码或工具执行能力，各家通用助手的产品差异越来越小。
独立产品需要重新获取用户、培养使用习惯。已有助手增加功能，则可以沿用账号、入口和用户关系。对用户来说，一项完整任务也常常跨越多个方向——先搜索资料，再读取文件，最后制作报告或PPT，能够连续完成这些步骤，可能比拥有多个互不相通的工具更有吸引力。
只是，当越来越多通用助手都能搜索、写作和生成内容，差异就不再只由功能清单决定。结果是否可靠、任务能否连续完成、需要用户反复纠正多少次，以及服务价格是否合理，都会影响用户最终留下来用哪一款。
浪潮翻涌了四年，十大赛道已经挤满了产品。QClaw的188天退场并非孤例，而是这场残酷竞争的常态。旧产品接连退场或走向整合，而各家关于下一代超级入口的较量，才刚刚开始。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The mirror studio, by Huang Ying, editor of Qi Humiao
A lobster is ready to leave, and the next round of competition for a personal assistant has begun.
On 24 September, TencentClaw announced that, as a result of business development adjustments, it would cease operations on 24 December 2026 and provide tools to help users migrate data to WorldBudy, another product of Tencent. QClaw was only 188 days away from 20 March when it was officially detected that the suspension had been declared.
On the other side, the major factory is in the middle of a new track. On 13 October, Ali thousands asked about the forthcoming publish personal assistant product, “Mo!” and tried to help users deal with delays in processing airline claims, shopping prices, telephone appointments, etc. Previously, the byte beat was also being exposed to the promotion of the personal assistant product, with the project codenamed “Spell”.
This has been repeated over the past four years. We combed Tencent, Alibaba, Baidu, byte, and the dark side of the moon, Zhipu, Mini Max, DeepSeek, eight companies, on 10 main AI tracks.
The data present two core features: first, the extremely short cycle of competition, from text-to-image, ChatBot to video generation, AI Agent, a continuous stream of new winds; and second, the shorter life cycle of the product, with a change of name, integration and stoppage, with stand-alone products increasingly coming closer to the General Assistant and the bottom platform.
The company continues to broaden what AI can do and to correct what is worth doing as a single product.
Four years of 10 races: more windways and faster follow-up in large factories
Over the past four years, the proliferation of iconic AI products has often led to a redistribution of attention across the industry, and these changes have left a clear mark on the timeline.
In 2022, text-to-image tools, such as Midjourney, Stable Diffusion and others, received attention, “input a sentence, generate a picture” becoming the starting point for many people to recognize the generation of AI. Domestic enterprises quickly followed up with the Ernie I of Baidu and the Qwen equivalent of Ali, and AI officially entered drawing, design and content creation.
The large-language model ChatGPT then pushed the tide to a wider audience. In 2023, products such as Ernie for Baidu, Qwen for Ali, Doubao for Z113Z, and Kimi for the dark side of the moon became more visible. The chat box is gradually becoming the basic form of the AI application, around which different needs, such as search, office, escort, are also built.
By 2024, the industry had explored in two directions: on the one hand, multi-modular video generation, Jimeng of byte, quick-hand Kling of Zhipu of Zhipu of Z, and on the other hand, Al hardware, Ari, Baidu, Z113Z, etc., tried to place smart assistants in headphones, Ai glasses, etc., to secure user access outside the App mobile phone.
The first wave of Ai's main track is becoming shorter and faster for companies.
In January-March 2026, for example, a variety of lobsters followed. In March alone, there were byte, Tencent, Zhipu, Baidu, and Ali five companies, which concentrated on their own “Claw category” products. Compared with the long months of early follow-up from ChatGPTpublish to LLM in the country, new hot spots are now leaving little time for companies to watch.
These waves are not simply “substituted in turn”. While new directions are emerging, investment in the old directions continues. Rather than looking for the next windfall, companies are faced with the need to decide in a cumulative layout which products are worth continuing, which capabilities can be combined and which tests should end.
"All for the big factory." Start-up company, "Priority."
Of the 10 tracks in this count, Alibaba, byte beats cover all 10 and Tencent, Baidu each covers nine. In contrast, the dark side of the month covered six, Zhipu five, Mini Max four and DeepSeek only two of the four companies.
Behind these differences are differences in corporate resource endowments and business priorities. For large Internet factories, AI is not only a tool for opening new businesses, but also for remodeling old ecology. Whether search, content, office, electrician or hardware, the vast existing user and business scenes provide them with enough test fields and more markets to hold.
It is even more important for AI to make strategic trade-offs: decide where to prioritize core resources.
What products remain on the field after the track has been opened?
Access to a track does not mean that an independent product can be maintained on a continuous basis. As can be seen from the statistical chart, after the opening of the track, families are constantly adjusting the shape and retention of their products.
The base entrance is fully maintained, but other tracks are frequently shuffled.
Eight companies have all deployed chat assistants, which are the basic entry points for carrying out search, creation and task handling. In other directions, however, there are cases where the product exits and stops.
For example, in the past four years, Mini Max's Glow has fallen, Ohai on the dark side of the moon has stopped, Tencent's music has stopped, and Baidu's monthly box has stopped. However, within the same track, products such as Talkie, Starfield and Cat Boxes are still functioning.
After wide distribution, large plants generally move towards internal integration and convergence.
byte is characterized by a wide layout, while constantly adjusting the division of labour among products. Doubao, Jimeng, Cat Box, each carrying different demands, the star drawing team is merged with Doubao, the MarsCode programming assistant is changed to TRAE plugin, and the TRAE team is later integrated into Doubao...
Some of Baidu's capabilities converge with those already in place. Ernie was moved to Ernie; Ernie fast code was incorporated into Baidu; and chat assistants experienced a “Ernie” change of name and entry. It continues to engage in multiple directions while reducing partial access to independence.
Ali and Tencent have the same integration. Ali's Goku was integrated in July 2026; Tencent's QClaw has announced that it will be suspended in December 2026 and that it will direct users to the WorldBudy.
The path to start-ups is divided.
DeepSeek ' s consumer product layout is relatively concentrated. Compared to other companies, which continue to introduce new names and new entrances, it focuses on increasing the capabilities of smart assistants. Zhipu's text-to-video product, “Smart” was incorporated into “Zhipu Statement”; MiniMax's general assistant, “Hailuo AI”, was later renamed MiniMax, with the original brand shifting to video business.
The suspension or consolidation of products involves not only changes in user demand, increased competition in the market, but also intra-firm resource trade-offs. The layout of publish is a starting point, and it is only after continuous input that the company can decide which products deserve to exist independently.
Tracks of six typical products: name change, business separation and consolidation
An iterative trajectory of six typical products is observed, with one recurrent move: a name change.
Ernie of Baidu was replaced by a textual statement and later by Ernie; and Z6Z of Ali was simplified to Qwen and then replaced by a question of thousands. From product opening to a second change of name, the two took 1029 and 793 days respectively.
The change in name is often accompanied by a reorganization of the company ' s product positioning and operational boundaries.
After about 10 months on line, Minimax' Hailuo AI's General Assistant was renamed MiniMax, while its original “Hailuo AI” brand was turned to video business. The name still exists, but the main purpose of the reference has changed.
The MarsCode programming assistant for Z113Z changed his name to TRAE after about 10 months in the country. The original programming assistant was placed in a larger product system as one of the means of use with the launch of the TRAE independent development environment.
Quarque continues to be adapted along the original product. From July 2024, when the Super Search Box was upgraded, to March 2025, when the publish “AI Super Box” was upgraded, 246 days passed. It attempts to expand user-friendly search portals to AI applications capable of handling more tasks.
There are, of course, opportunities for products to “rename”. QClaw of Tencent opened a public survey in March 2026 and announced a cessation of service (scheduled to cease service in December) in September, only 188 days after the announcement. In less than half a year of life, the error rate in refractions of products tested on new tracks is decreasing dramatically.
For companies, integration can concentrate inputs and reduce duplication; for users, frequent name changes and migration also mean re-identification of access and adaptation to operational practices.
The General Assistant is becoming more and more like a functional hybrid.
Independent products are increasingly fragmented and integrated, and more and more can be done within the general assistant.
From chatting, to online search, document understanding, to drawing, voice, in-depth research, document generation and implementation codes, which were discussed separately, the same assistant was gradually installed.
The public records of the four sections sample show that by the end of 2024, Doubao, Quiz, DeepSeek and Kimi had shared three capabilities: chatting, online search, and documentation understanding. Since then, research, office and implementation capacity has continued to move into some of the assistants. By October 2026, Doubao, Quiz, and Kimi had online documentation generation and code or tool implementation capabilities, and the products of the general assistants were becoming smaller and smaller.
Independent products require re-acquisition of users and development of usage habits. With the helper added, account numbers, access points and user relationships can be maintained. For users, a complete task also often cuts across multiple directions — searching for information first, reading documents and finally producing reports or PPTs — and is likely to be more attractive than having multiple disparate tools to complete these steps.
Only when more and more general assistants can search for, write and generate content, differences are no longer determined solely by functional lists. The reliability of the outcome, the continuity of the task, the number of times the user is required to correct it repeatedly and the reasonable price of the service will affect which paragraph the user will ultimately retain.
After four years of waves, the 10 main tracks are full of products. The 188-day exit of QClaw is not an isolated case, but a normal feature of this brutal competition. The old products are coming back or moving towards integration, while the competition for the next generation of super portals is just beginning.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.