In-depth Report
-
Tencent Marvis (nicknamed "Mavis") is an operating system-level personal AI assistant launched by the Tencent Yongbao team on May 20, 2026. It is not just a chatbot, but turns the entire computer into a conversational "intelligent object", directly accessing the bottom layer of the system to achieve closed-loop control from file management, system settings to cross-end collaboration. The product adopts a collaborative architecture of "1 main Agent + 5 deputy Agents", providing 10 million free tokens per person per day, supporting two working methods: efficiency mode (device-cloud collaboration) and privacy mode (completely local), and is targeted at ordinary users and professionals.
-
Marvis was developed by the Tencent Yongbao team, inheriting the more than ten years of cooperative relationships between Yongyongbao and software and hardware manufacturers and four years of cross-end engine technology. At the level of technical cooperation, Marvis has established strategic cooperation with Microsoft and Intel respectively. Based on the Microsoft Foundry on Windows end-side AI development platform, the WinML inference framework enables the optimized operation of large models on local NPUs, GPUs and CPUs. Marvis will begin internal testing (invitation-based) in April 2026, and will be officially fully open on May 20, supporting Windows, macOS and Android. The iOS version will be launched in mid-June. The product does not require an invitation code, and users can directly visit the official website marvis.qq.com to download and use it. AI assistants currently on the market are roughly divided into three categories: general conversational (such as Doubao, Kimi), tool-based (such as AI PPT, programming assistant) and operating system level (such as Microsoft Copilot, Apple Apple Intelligence). Marvis clearly chose a Category III track.
-
The core of Marvis is a collaboration system of one main Agent and five deputy Agents. As the "commander", the main Agent is responsible for understanding user intentions, breaking down task steps, scheduling the execution of corresponding Agents, and summarizing the results. The five deputy agents perform their own duties: the File Agent is responsible for file management and search, the Computer Agent is responsible for system operation and hardware detection, the App Agent is responsible for application control, the Browser Agent is responsible for web page interaction and data capture, and the Search Agent is responsible for network search and information aggregation. Marvis designed a quasi-physical workspace for these Agents - there is a gym, a lounge, and each Agent has its own work station. In their free time, you can see some of them drinking coffee and some exercising. This design is quite interesting. Marvis integrates end systems, files, applications, computing power, and cross-end connections into the same AI middle layer. Users don’t have to remember to set up entry points, file paths, or software operation steps; they only need to describe the goal in one sentence. In actual experience, it is easy to check computer configuration, battery health status or network status in one sentence. Use it to check whether the hardware can run a certain game smoothly. It can automatically read the configuration and compare the minimum and recommended requirements to give a conclusion. When it comes to file searching, Marvis does a pretty good job. Users can use natural language to describe file characteristics, such as "Find the photos I took at the Great Wall" or "Help me find the reimbursement form that was modified last week." It not only does keyword matching, but also understands time range and semantics. The image search is even more of a highlight - let it find screenshots of a certain game character, and it will take about 3 minutes to find it accurately, including images of the characters in group photos.
-
Marvis is currently completely free, no invitation code is required, and each person has 10 million free tokens per day, which can meet the daily use needs of the vast majority of individual users. Based on regular usage, this amount can support thousands of conversations with Marvis per day. However, the relevant person in charge of the Tencent Marvis team has publicly stated: "It is unrealistic to be completely free in the long term." The computing power cost of AI services is relatively high, especially the call of large cloud models requires continuous investment. In the future, a model of "free quota + fee for excess amount" or "free basic functions + fee for advanced functions" may be launched.
-
Based on multiple evaluation platforms and user feedback, positive reviews of Marvis mainly focus on several aspects. For computer novices and users with elder care needs, Marvis has greatly lowered the threshold for computer use. One user shared his father's experience - his father accidentally turned off the display of desktop icons, thinking it was a virus on the computer. He solved the problem in a few minutes by remotely controlling Marvis. The improvement in daily efficiency is also obvious. According to user statistics, using Marvis saves about 2.5 hours per week - time that was originally spent on trivial matters such as finding files, adjusting settings, and cleaning out trash. Finding files is shortened from 2-5 minutes to 10 seconds, and organizing the desktop is shortened from 15 minutes to 2 minutes. Negative feedback mainly focuses on several aspects. Although the image search function is outstanding, text keyword searches sometimes have matching logic problems. Some reviews mentioned that when searching for "Tumu Fort", the results listed more than 100 pieces of content, including dozens of pure English documents. The degree of automation of the computer housekeeper function is not enough. It takes nearly 20 minutes to clean the C drive and finally outputs an analysis report for you to operate by yourself.
-
Industry media generally believe that Marvis represents the trend of AI assistants moving from “cloud conversation” to “terminal execution”. Mainstream media such as CCTV, Global Network, and China Science News have all given positive reports, focusing on its operating system level positioning and cross-end collaboration capabilities. An industry analyst said in an interview with Global Network: "The AI assistant of the future must be multi-device and cross-device. Users will not only use AI on one device. Mobile phones, computers, tablets and even car systems need to be seamlessly connected." The Tencent Cloud Developer Community has launched a series of Marvis usage tutorials, covering a wide range of topics from installation and configuration to knowledge management to automated workflow.
-
This is the biggest concern for Marvis right now. There are authorization options for image recognition and document recognition in the system, but some users reported that without clicking authorization, the authorization interface disappeared directly when they opened it again the next day, and all applications, documents and pictures in the computer had been scanned and classified. There is currently no way to cancel authorization. The cloud efficiency mode requires data to be transferred to the cloud for processing by default. Although officials state that data is transmitted encrypted and not stored persistently in the cloud, this transparency is far from enough for privacy-sensitive users. Privacy mode can theoretically solve this problem - it is equipped with a large end-side model, runs completely locally, and can be used when disconnected from the network. However, this mode has extremely high hardware requirements (starting with 16-core CPU + 32GB memory), and many users cannot even see the enable button. This means that most users can actually only use Productivity Mode.
-
Marvis is most suitable for the following groups of people: first, computer novices and elder users, who do not need to remember complex operation paths; second, professionals, who need to frequently process files, organize information, and adjust system configurations on a daily basis; third, users who need to work collaboratively across multiple devices. People who are not suitable for it include: users who are extremely sensitive to data privacy and need to handle highly confidential information (the hardware threshold of privacy mode is too high); and users who expect it to become a "fully automatic magic wand".
-
Tencent Marvis is an important step for AI assistants to move from “dialogue” to “execution”. It turns the entire computer into a conversational intelligence, making natural language control systems a reality. The free strategy lowers the threshold for use, and the daily quota of 10 million Tokens is more than enough for most users. At present, there is still a gap between the grand positioning of "operating system level AI assistant", but as a free product, it is already a very practical daily efficiency tool and is worth a try.
User Reviews
-
CherylEdwards—用了两周 Marvis，最惊艳的是那个图片搜索功能。我说「找一张去年在海边拍的照片」，它真的一点点帮我扫了整个图库。不过 Token 烧得也太快了，搜了一轮一百多张图就烧了快两百万，一天一千万看起来多，图片玩多了还真不够用。 -
胡哲贞—帮老爸装了一个 Marvis，他电脑白痴一个，以前出点问题就得打电话让我远程弄。现在直接跟 Marvis 说就行了，前几天他自己用 Marvis 搞定了打印机驱动，把我惊到了哈哈。 -
realChakiraVan der zouwen_2024—隐私模式确实是个好东西，但门槛也太高了吧。我的笔记本 16G 内存直接灰了按钮，只能用效率模式。问题是效率模式默认要把数据传云端，虽然官方说加密，但心里还是有点不踏实。 -
梁军—「帮我把桌面整理一下」这个指令我每天都要说一遍，它自己会把文件按类型归类，还会问我要不要删重复的。以前我桌面跟垃圾堆一样，现在终于干净了，真香。 -
EmilyGomez351—说一个大家可能没注意到的细节，Marvis 清 C 盘是真的分级处理的。哪些可以放心删、哪些要确认后才删、哪些绝对不能动，分得清清楚楚。我之前用各种清理软件一键清空，经常清完发现什么功能不对了，用 Marvis 还没出过问题。 -
Adam_Morgan_88—手机远程控制电脑好好用！！人在外面突然要发个文件，掏出手机打开 Marvis 直接找到文件发微信，全程一分钟。以前还得找台电脑或者让同事帮忙，太麻烦了。 -
Nicole986—那个 Agent 工位的动画我太喜欢了哈哈哈，每个小牛马都有自己的工位，有时候看它们在摸鱼打游戏就觉得好笑。虽然没啥实际作用，但是情绪价值拉满。 -
8abd73qupx—深度用下来发现一个坑：有些 App 它操作不了。我想让它帮我在微信上发个消息，被拒绝了，小红书也不行。说是安全原因，但这样一来实用性就打了个折扣。 -
Abigail_StewartX621—Marvis 跟我之前用的 OpenClaw 最大的区别就是——不用折腾。装好就能用，不需要自己搭工作流配 skill，出门左转那些 Agent 对普通人来说真的太难上手了。 -
DorisButler520—每天一千万 Token 是真的够用，我上班重度使用一天也就消耗两三百万。但我听做设计的同事说，他要处理大量图片识别任务的时候一天能用掉八九百万，所以看人吧。 -
Jack.Brown_202083—清理自启动项这个功能太实用了。让 Marvis 检查一遍，它把微信、钉钉这些必用的给保留了，把 OneDrive、Steam、Adobe 这些偶尔用的建议关掉。关了之后开机速度快了一截，从以前四十多秒降到二十几秒。 -
Kathryn.Perry_7—唯一不太满意的是那个本地知识库的扫描权限问题。我还没点授权呢，第二天打开发现它已经把我全盘扫描完了，分类确实准确，但这个不经过我同意就扫描的行为让我有点不舒服。 -
潘兰—Marvis 的搜索是真的强，不像 Windows 自带那个垃圾搜索只能搜文件名。我跟它说「帮我找上个月关于合同模板的 PDF」，它居然真的找到了，而且是基于内容匹配的，不是靠文件名。 -
MHarris_88—查 Wi-Fi 密码这种小功能我居然用了好多次。家里来朋友问密码，直接一句「帮我查一下 Wi-Fi 密码」就搞定了。朋友都惊了，哈哈哈。 -
M_aria53—说实话现在的 Marvis 还没有杀手级功能，单看每一项都不算惊艳。但组合起来确实能省不少时间，我统计过一周大概省了两个小时，主要是找文件、清垃圾、调设置这些琐碎事上。 -
Catherine.White_88—跨端协同的远程操控延迟很低，比我用过的向日葵和 TeamViewer 都流畅。我人在外面，手机远程操控家里的电脑剪视频，虽然操作精度不如直接坐电脑前，但应急完全够用。 -
RoyMorrisZ028—让 Marvis 帮我查电脑能不能跑黑神话悟空，它自己读了配置然后去网上对比，最后告诉我显卡刚好过最低线，内存够用，建议降低画质。比我一个个参数去比对方便太多了。 -
Gregory_WhiteSr—今天试了让 Marvis 把二十张 Excel 汇总成一张统计表，中间问了我好几轮问题，最后出来的结果还需要我手动调整。感觉复杂任务上还有很长一段路要走，简单的系统操作倒是很稳。 -
goldenkoala824—免费是真的香，每天一千万 Token，对于我这种轻度用户来说根本用不完。而且不像 ChatGPT 还要折腾网络问题，国产软件用起来就是省心。 -
AustinBarnesK—我就想问一个问题，本地模式能不能降低一点门槛啊？看了下要求 16 核 CPU + 32G 内存，我这台去年买的办公本直接不配拥有隐私安全了呗？希望能出个轻量版本地方案。