Tencent has created a "Jarvis": It can sign in and change configurations for me, and it even takes naps and goes to the toilet. Here's a hands-on test.

According to a report by Zhidongxi on May 15th, recently, Tencent has started the internal testing of an operating - system - level personal AI assistant named Marvis. This AI assistant completes App operations, EXE operations, computer operations, file management, document generation, and various complex tasks through the collaboration of multiple Agents. It is online 24/7 and supports cross - device operations.

Currently, Marvis only supports Windows PCs and Android phones. iOS/macOS support will be launched soon.

In daily usage scenarios, Marvis can handle quite a few tasks. In the work scenario, it can perform file format conversion, contract information review, operational data analysis, and even help organize references. These functions go beyond simple text and code generation and can truly operate more underlying elements such as the operating system, files, and applications.

In the life and entertainment scenario, it can help you monitor limited - time game benefits and celebrity news, automatically complete Weibo check - ins, organize high - definition materials, and also perform daily morning news summaries on a scheduled basis.

The name Marvis pays tribute to the well - known intelligent assistant "Jarvis" and is a combination of "Ma" and "Jarvis". Interestingly, it designs the image of each Agent as a black pony with horns, simply referred to as a "little pony". The "little pony" also inherits Tencent Penguin's iconic scarf.

You can check what each Agent is doing at any time in the virtual office. Agents assigned tasks will sit at their workstations and work hard, while other idle Agents may doze off, wander around, work out, drink coffee, or go to the toilet.

When you assign a task, an Agent supervisor will come out to take charge and assign tasks to the corresponding Agents. After completing the work, the Agents will report to the supervisor.

Currently, the expert Agents in Marvis have the capabilities of App operation, computer system maintenance, web interaction, digital asset management, and search tasks.

Marvis currently offers two operation modes: the cloud - based efficiency mode uses the latest models such as Hunyuan and DeepSeek V4, while the local privacy mode uses the Qwen edge - side model, ensuring that no data leaves the computer.

This product also supports the function of a local knowledge base, which can automatically identify and index local documents and pictures. When users need to search for specific content, Marvis can perform title - based or content - based matching.

Compared with other "lobster - like" products that can also control computers, what are Marvis's advantages?

Zhidongxi learned that Marvis is developed by Tencent's Yingyongbao team, which inherits more than a decade of cooperation relationships with hardware and software manufacturers such as Intel and Microsoft and cross - device engine technology.

Therefore, Marvis can not only operate computer EXE software but also directly control mobile phone Apps on the computer through Yingyongbao (Apps such as Flush, KaiPanLa, Feichangzhun, Autohome, and Vipshop have authorized it).

The team has optimized the edge - side model at the chip level, achieving a 20% increase in the running speed of the edge - side model and enabling direct query and modification operations on computer files through prompts. Most current AI applications do not have such in - depth operating - system - level capabilities.

Zhidongxi obtained the internal testing qualification for Marvis as soon as possible. Based on our actual experience, Marvis shows the prototype of an "AI worker". It can truly help you with tasks, and the quality of task completion is good, but there is still room for improvement in terms of speed, token consumption, and flexibility.

For more information: marvis.qq.com

Find files and adjust settings

Marvis has tried it all

Computer operation is one of Marvis's strong suits. Marvis has system - level capabilities and can directly complete the entire process of computer operations. Our first - batch cases revolve around such tasks.

After entering the homepage, users can directly talk to Marvis and assign tasks, or let Marvis perform scheduled tasks through the automatic task page.

Case 1: Capable of image search and file location, but a bit token - consuming

For many users, finding a specified file or image on a local device is a time - consuming and laborious task. I tried to ask Marvis to find a picture containing NASDAQ elements in my local download folder.

After receiving this task, the team leader Agent of Marvis assigned the task to the file - processing Agent. The latter first tried to search by file name but found no match. During this process, it judged that most of my picture names were meaningless and decided to change the approach, using image analysis to find the relevant image.

After changing the approach, Marvis finally got on the right track and successfully found the picture containing NASDAQ - related elements. However, when using the image analysis tool, Marvis's token consumption increased exponentially. After searching more than a hundred pictures, the total consumption reached 2 million tokens.

Marvis can complete cross - device file searching and transfer. I sent a command to Marvis on my phone to search for files related to ByteDance's valuation and send them to my phone.

It quickly found the files and sent them to my phone. Maybe because the file names themselves contained the keywords to be searched, the token consumption of this task was significantly reduced.

Case 2: Accurately capture computer configuration and give optimization suggestions

Marvis can directly perform some operations at the Windows system level. When I asked it to change my computer's theme to green, Marvis's computer operation Agent took over the work and directly found the setting entry for the system theme color. This operation was not achieved through simulated clicks, indicating that it delved into the system's underlying layer.

However, it did not complete the task 100% and stopped at the step of finding the entry. The subsequent specific settings needed to be done manually by me.

Marvis can accurately obtain various system information. When I asked it whether my device could support large - scale games like "Black Myth: Wukong", it quickly sorted out my system configuration, seemingly directly finding the relevant content in the settings.

Then, I asked Marvis to analyze the weaknesses in my configuration and give upgrade suggestions. It judged that my graphics card configuration was sufficient, but the CPU, memory, and hard drive were a bit outdated. Marvis suggested that I upgrade the memory, saying that it was the most direct and cost - effective way to improve performance.

However, it seems not to be aware of the current situation of rising memory prices and told me that a 32GB DDR4 3200MHz memory only costs four or five hundred yuan. In fact, the price of similar memory is at least over 1000 yuan.

In our actual experience, Marvis will definitely ask users when encountering tasks such as batch file deletion and core configuration modification, and will only execute after the user clicks to confirm.

Write code and generate reports

Multi - Agent collaboration has highlights

In theory, the collaboration of multiple Agents can complete more complex tasks or improve the quality of task execution. We tested Marvis's ability boundaries with several complex tasks.

Case 1: Develop a local knowledge base without touching code

Our first complex task required Marvis to design a local knowledge management system that supports custom large - model access and realizes simple RAG.

In this more complex task, Marvis did not call multiple Agents but had the main Agent complete all the development work. Marvis successively created the directory structure, wrote the configuration file and core modules in parallel, and then built the Web interface. Marvis defaulted to using the ollama local running model, and I asked it to switch to the mode of accessing an external API.

The API setting interface provided by Marvis is quite user - friendly. It sent me tabs, and I just needed to select and fill in the information step by step.

Finally, Marvis built this knowledge management system according to my requirements, supporting options such as importing files, importing directories, and importing text, and providing reference sources. We tried uploading several files and asking relevant information, and the content output by the system was accurate and did not fabricate information not in the knowledge base.

Case 2: Summarize the Silicon Valley lay - off wave and directly output a PDF file

The next actual - test task required Marvis to sort out the lay - off situation of large Silicon Valley companies this year and create a PDF webpage containing lay - off data and analysis, with a separate source of information attached to each piece of data.

In this task, Marvis used the multi - Agent collaboration mode. The file Agent was activated and actively called the document - writing Skill. This Skill requires the model to first judge the complexity, then load the writing template and writing specifications, and then start the formal execution.

Marvis judged that this was a complex document, expecting an output of more than 3500 words. It conducted multiple rounds of web searches, and after collecting enough data, it started writing the PDF content page by page, finally outputting a complete research report.

During the execution process, Marvis's multi - Agent collaboration made the task division more reasonable, and the parallel work of two Agents improved the efficiency to a certain extent. However, in this task, the Agents had the problem of automatically switching to English, which may be caused by the content of the Skill and the characteristics of the underlying model.

The document generated by Marvis can be previewed directly in the sidebar. From the results, the content of this document is significantly more detailed than that generated by a single AI, and the data in the text is basically accurate.