On June 23, the iOS version of Marvis, Tencent's AI assistant at the operating system level, officially launched on the Apple App Store. This marks the completion of full coverage for both mobile and PC platforms, following the release of the Android, Windows, and macOS versions on May 20. Users can now summon this "24/7 online" AI assistant anytime, anywhere, and seamlessly switch between their phone and computer, integrating AI into every scenario of daily work and life.

image.png

The interface of the iOS version continues Marvis' consistent design language: upon entering the app, the main screen displays the conversation entry "Hello, what would you like to do today?" and preloads high-frequency scene quick commands such as "Bangkok Travel Guide Webpage," "Stock Information Inquiry on Tonghuashun," and "Personal File Organization," lowering the user's initial learning threshold.

iOS Version Features: A Portable Desktop Terminal on Mobile

The launch of the iOS version focuses on enhancing the collaborative experience between mobile and desktop devices. Through "multi-device account synchronization," users can connect and control multiple devices with just one phone, whether an Android phone connecting to a Mac or an iOS phone connecting to a Windows computer; using the "cross-device remote control" capability, even when not near the computer, users can take out their phone and remotely control the computer desktop in real time through Marvis to complete complex tasks that were previously only possible on a PC, while also being able to see the computer screen in real time, achieving true "desktop-level visual remote control," rather than traditional command-based remote commands.

image.png

For example, in a daily work scenario, if a user needs to supplement a document while commuting home, they can simply ask Marvis to send the meeting minutes from the computer desktop to the mobile app, and Marvis will automatically complete cross-device positioning and transmission; when abroad and needing urgent historical documents, they can also remotely retrieve, export, and send them to their email. For users who need to maintain computers for their elders, they can also remotely take over the home computer via their phone to complete software installation and configuration adjustments, and see the actual situation of the computer, avoiding the trouble of describing the problem over the phone.

In addition, even when the computer is locked, users can still use their phone to remotely enter the password and perform subsequent operations. For time-consuming tasks such as game updates or large file downloads, users can check the progress and issue subsequent instructions at any time via their phone. Even when the computer is offline, the iOS version of Marvis can switch to a pure cloud mode, independently completing tasks such as information retrieval and file processing.

Local + Cloud Dual Mode: Exploring a New Form of Operating System-Level AI

Differing from most AI assistants that run on the "application layer," Tencent's Marvis takes a "system-level" technical approach and bets on a "local mode + hybrid architecture of end and cloud." The product offers two options: "Efficiency Mode" which balances speed and accuracy through end-cloud collaboration, and "Local Mode" which runs based on the local large model on the device, ensuring that files do not need to be uploaded to the cloud, maximizing the protection of user data privacy.

This choice addresses current industry pain points of AI agents: mainstream AI agents generally rely on pure cloud operation, leading to four major bottlenecks—high costs, privacy concerns, response delays, and unavailability when offline. Each call consumes cloud computing power and Token, sensitive files must be uploaded to the cloud, network fluctuations slow down responses, and offline means "no resources available."

Marvis' solution involves deep strategic cooperation at the chip and system levels, allowing local and cloud to perform their respective roles. On the technical collaboration front, Marvis has optimized chips with Intel based on OpenVINO, and the team has completed extensive device compatibility work, providing hardware-level support for local models to run smoothly and stably. At the system level, Marvis has achieved system-level file and setting operation permissions with Microsoft. This path, from bottom-up capability accumulation to system-level ecosystem collaboration, forms the core competitiveness of Marvis, distinguishing it from general cloud-based agents.

On the other hand, unlike other AI products that require uploading to the cloud for processing, Marvis comes with a small on-device model that can run locally, performing indexing and preliminary analysis on the user's computer. Even in efficiency mode, after the user authorizes the selection range, all recognition, organization, and search of documents and images in the local knowledge base are completed locally on the device, without leaving the computer, ensuring both privacy security and significantly reducing cloud Token consumption.

Currently, Marvis is still in a period of rapid iteration of its capabilities. Since its launch on May 20, 2026, Marvis has maintained a high frequency of iterations, delivering 21 versions within just one month, averaging one version every 1-2 days. These include features such as session management, voice input, connecting to Mac for remote control on the phone, user profile management, and low-resource mode. The team stated that in the future, they will gradually introduce new features such as the Mac local mode, global knowledge base, and custom agents, continuously improving the product experience while involving users in the growth process of Marvis.