
on may 20, tencent officially launched its brand-new operating-system-level ai assistant—marvis (ma weisi)—and simultaneously opened the official website (marvis.qq.com) for download. it’s available to install and use right away with no barriers or invitation codes required. unlike traditional conversational ais or single-point application-layer agents, marvis is deeply embedded in the terminal system’s underlying layers. with “native os intelligence” as its design foundation, it seamlessly connects five key dimensions—system, files, applications, computing power, and cross-device connectivity—to build a truly understandable, schedulable, executable, and collaborative personal computing hub.
more than just conversation—it’s an intelligent entity that takes over your entire computer
for the first time, marvis elevates the pc from a “passive tool” to an “intelligent object capable of natural interaction.” it doesn’t rely on specific entry points or memorized paths; users simply issue a spoken command like, “help me check if my graphics card can run ‘black myth: wukong,’” or “sort all pdfs with invoice text from last month by date and send them to my email,” and the system automatically parses the intent, breaks down the steps, invokes the corresponding capability modules, and proactively hands back control at critical junctures—truly achieving “speaking human language, handling human tasks, and respecting boundaries.”
system-level understanding makes setup, diagnostics, and optimization accessible with one click
by leveraging deep modeling of windows kernel mechanisms, hardware status, service processes, and registry structures, marvis can respond in real-time to system-level queries and operations: checking battery health, identifying slow boot causes, disabling redundant startup items, clearing temporary caches, adjusting power policies, and more. all actions are executed through semantic-driven commands, eliminating the need to open control panel, type commands in the command line, or consult manuals. for beginners and senior users, this provides an ultra-simplified gateway to digital life; for office workers, it serves as a one-stop efficiency center that aggregates capabilities across multiple tools.
ai-powered personal knowledge base turns “can’t find the file” into a thing of the past
marvis comes equipped with multi-modal local models that support cross-format content understanding: not only can it search by file name, but it also recognizes text in images, extracts core passages from documents, identifies facial features, detects holiday scenes, or recognizes geographic tags. based on this, it can automatically generate personalized knowledge spaces such as ai image libraries and ai document repositories, transforming scattered digital assets into searchable, interconnected, and reusable knowledge nodes, thoroughly resolving the common pain point of “remembering the content but forgetting the file name.”
security isn’t optional—it’s built-in by default
marvis employs an l2-level dynamic security safeguard mechanism: any high-risk operation involving file deletion, system configuration changes, or permission modifications must first trigger a “hard inquiry” process—first generating a readable execution plan, then waiting for explicit user confirmation. payment-related actions are completely isolated on the device side, requiring users to personally complete verification. security logic isn’t an afterthought—it’s embedded throughout the entire workflow, from task analysis and planning to execution.
no compromise between edge and cloud: the dual optimal solution for efficiency and privacy
marvis offers a dual-mode operating mechanism:
- efficiency mode: complex intent understanding and long-term planning are handled by cloud-based large models, while file parsing, image recognition, local indexing, and action execution remain entirely on the device, balancing response speed with data control;
- privacy mode: entirely offline throughout the process, with all data staying within the device. model inference, ocr recognition, and dialogue processing are all performed locally, enabling offline use and meeting stringent compliance requirements in sectors such as finance, law, and hr.
its collaborative logic isn’t simple traffic splitting—it’s layered offloading: the device performs lightweight preprocessing first (such as image cropping, text chunking, and metadata extraction), uploading only essential abstract information to the cloud, significantly reducing token consumption and minimizing data exposure.
out-of-the-box ai collaboration team: six specialized agents working in parallel
marvis ships pre-integrated with six vertical-domain agents, forming a 24/7 online ai workgroup:
- main agent: orchestrates the overall workflow, understands raw user requests, and intelligently schedules tasks;
- file agent: manages digital assets, supporting full-text search, multi-format editing, batch conversion, and semantic archiving;
- computer agent: a systems maintenance expert, covering hardware diagnostics, driver updates, performance tuning, and troubleshooting;
- app agent: a desktop application scheduler, capable of launching, operating, and even automating local exe programs and third-party software;
- browser agent: a web interaction engine, supporting automatic login, form filling, dynamic page scraping, and structured data extraction;
- search agent: a global information aggregator, skilled at cross-validation, source attribution, and summary generation.
a user can issue a complex command in a single sentence, and multiple agents can respond in parallel. the execution process is transparent and visible, with results automatically aggregated into a unified output area.
seamless cross-device collaboration—your phone becomes an extended display for your computer
marvis natively supports a unified account system across windows, macos, ios, and android, enabling seamless interoperability and state synchronization between devices. it not only allows you to remotely view your pc desktop and take over input in real time from your phone, but also lets you enter passwords and unlock your pc while it’s locked—all directly from your phone. high-frequency tasks such as web browsing, file transfers, and screenshot annotation can all be smoothly continued across devices. unlike the imperative interaction of traditional remote-control tools, marvis delivers a desktop-level collaborative experience with contextual awareness and visual feedback.
currently, each user enjoys a daily free token allowance of 10 million. with the launch of the ios version, integration of the mcp protocol, expansion of app permissions, and continuous upgrades to on-device models, marvis is accelerating the transformation of ai agent capabilities—once exclusive to tech enthusiasts—into everyday smart companions that are accessible, reliable, and trustworthy for the general public.