Rabbit Launches OS3 Without R1, Bets Comeback on Cloud Autonomous OS
Rabbit unveils autonomous OS3 that works without R1, running on Windows/Mac/Linux with 5-device support and model choice via cloud.
Rabbit has signaled a shift away from dependence on hardware. On Sept. 22, 2026, the company unveiled OS3, an autonomous AI agent that does not require its dedicated R1 device. According to reporting by Emma Roth of The Verge, the new platform is positioned as an “agentic operating system.” It is designed to run in the cloud while executing operations locally on Windows, Mac, and Linux devices.
The company previously drew attention with the palm-sized AI device R1 shown at CES 2024. It promoted operation automation through a Large Action Model, but reception after launch was harsh. OS3 can be seen as a software-centered pivot reflecting those lessons.
Aim of Launching an Autonomous OS Without R1
The biggest change in OS3 is that it does not assume the dedicated R1 hardware. According to reporting by Emma Roth of The Verge, Rabbit explained, citing earlier reporting by Wired, that OS3 is a standalone AI agent usable without R1.
The conventional R1 itself served as the execution environment for the AI agent. In OS3, the execution environment moves to the cloud, and users’ existing devices become the targets of operation. By eliminating the need to purchase dedicated hardware, the move can be seen as aiming to lower adoption barriers and expand the user base.
The company calls OS3 an “agentic operating system.” It emphasizes autonomous execution of multi-step work rather than single chat responses. The concept is to complete a series of tasks, including file operations and application integration, starting from user instructions.
Rabbit, the company behind the underwhelming R1 device, is rolling out a standalone AI agent that you don’t need its hardware to use, as reported earlier by Wired. The startup says its new OS3 “agentic operating system” runs in the cloud but operates locally across Windows, Mac, and Linux devices.
This description points to a hybrid structure combining inference in the cloud with execution on the device side. The core of processing remains in the cloud, while actual operations take place on the user’s devices. This enables access to materials stored on devices and to various applications.
Structure of the Execution Platform Spanning
Cloud and Devices
OS3 runs in the cloud, but operations are executed locally on Windows, Mac, and Linux devices. According to The Verge, the company describes this approach as “runs in the cloud but operates locally.” The cloud side handles planning and inference, while the device side handles concrete operations.
Users can register up to five devices under one account. OS3 distributes work across the registered devices and automatically selects the necessary machine. For example, the system is said to decide on assignments such as document creation on a laptop and image processing on a high-performance desktop.
Access routes are diverse. In addition to a dedicated desktop site, instructions can be given via messaging apps such as Telegram and iMessage. It remains usable from the conventional R1 device. The ability to issue instructions directly from everyday communication tools can be seen as a factor determining whether the agent takes hold.
When executing tasks, OS3 automatically identifies the necessary devices, files, applications, and AI models. Users do not need to specify the environment individually. It autonomously handles a series of flows, collecting information across multiple materials, calling appropriate applications, and integrating results.
This design aligns with the “cross-environment autonomous execution” pursued by recent AI agents. It is also highly compatible with the spread of development platforms premised on collaboration and instant sharing between agents, such as that reported in Agentgit, Git Platform for AI, Launches With Anonymous Instant Sharing. As OS3 spans devices and the cloud, the agent’s working area is no longer confined inside a single device.
How 5-Device Linking and Model Selection Work
In OS3, users can register their preferred AI models. The configuration is not fixed to a single model but allows switching by use case. Specific model names and the number supported have not been disclosed, but giving users a choice is significant.
OS3 automatically selects the optimal model according to task content. It is expected to call different models for work with different processing characteristics, such as document summarization, code generation, and image analysis. By having OS3 absorb switching between models, users are freed from the complexity of model selection.
Registration of up to five devices appears to be a limit assuming personal use of multiple devices. One agent bundles a distributed environment such as a work PC, a home PC, and a laptop. OS3 tracks where files are located and retrieves necessary materials from the appropriate device.
This mechanism blurs the boundary between cloud and device. Data remains on the device side, while control is handled by the cloud side. In exchange for convenience, users must delegate broad access rights to their devices. Permission design and transparency of operation history are likely to become evaluation criteria for enterprise adoption.
As shown by BAT Invests 3 Billion Dollars in Kling AI, Inside the Anti-ByteDance Coalition, competition over AI infrastructure is seeing simultaneous concentration of huge capital and decentralization. Rabbit’s approach of bundling multiple models without relying on a specific model can be seen as an option to avoid lock-in to a particular camp.
R1 Production Halt and Next-Generation
Cyberdeck Plan
Founder Jesse Lyu reportedly told Wired that the company has stopped manufacturing the R1 device. Reporting by The Verge also cites this point. It is a decision to stop new investment in R1 and concentrate resources on OS3.
At the same time, the company is not withdrawing entirely from the hardware business. It says it is developing a “cyberdeck” equipped with OS3 as its next device. According to Wired, release is expected “within months.” Detailed specifications and pricing have not been disclosed.
The term cyberdeck often refers to a portable development machine with a keyboard and display. Rabbit positions it for “vibe-coding.” Vibe-coding refers to a development method centered on natural-language instructions, generating code while conveying atmosphere and intent. Combined with OS3’s autonomous execution capability, it appears aimed at completing development work directly on the device.
R1 initially drew attention as a new form of AI hardware. However, response speed, limited functions, and insufficient differentiation from smartphones were pointed out. The production halt can be seen as reflecting the reality that single-purpose AI devices struggle to sustain in the market.
The shift to OS3 is a typical comeback strategy that compensates for hardware failure with software. By leveraging existing PCs and Macs, users can try the company’s agent without buying new hardware. It has the advantage of expanding the user base without bearing hardware inventory and distribution burdens.
Data Retention Explanation and Remaining Questions
Rabbit says it does not store, duplicate, use, or sell user data in OS3. At the same time, it says chat content and memories derived from it are retained on its servers. How these two explanations align will affect user understanding.
The reason for retaining chats and memories on servers appears to be context maintenance and personalization. Referring to past dialogues enables continuous work and reflection of preferences. However, the boundary between what is not stored and what is retained is not clear.
Furthermore, external AI providers linked to OS3 handle data in accordance with their own privacy policies. Separate from Rabbit’s policy, the policies of linked providers apply. Users need to check not only Rabbit’s policy but also those of their chosen model providers.
This structure carries challenges common to recent AI agents. The broader the access to files and applications on devices, the higher the convenience, but the wider the impact in case of leakage. Minimization of access rights and visualization of operation logs are required.
In wearables and small devices as well, data handling determines trust. As with health data management compared in Oura Ring 5 vs RingConn Gen 3 Full Comparison, business data and personal files handled by AI agents also form the basis of long-term trust. Rabbit’s explicit statement that it will not sell data is one declaration, but operational details such as retention periods, deletion methods, and third-party sharing remain in question.
OS3’s offering format, pricing, specific supported models, and availability of enterprise management features have not been disclosed at this point. Execution spanning cloud and devices presupposes a stable network and a resident agent on the device side. Operational details such as behavior offline and when devices are asleep require future verification.
Editorial Opinion
In the short term, the halt of R1 manufacturing and the shift to OS3 will be a litmus test for Rabbit’s business continuity. A method that leverages existing PC assets without relying on dedicated device sales significantly lowers adoption barriers. Over the next three to six months, actual task success rates, response speed, and the breadth of supported applications will determine continued use. How practically operation via messaging apps works will shape early reputation. In the long term, cross-device agents like OS3 could redefine the boundaries of operating systems and applications. On a one- to three-year view, the question is whether a portable autonomy layer not tied to a specific AI model or device will become standard for personal work environments. Freedom of five-device linking and model choice will also affect management of shadow AI and information control in enterprises. How to design the balance between convenience and control will be a condition for diffusion. The remaining question is whether Rabbit can demonstrate in a verifiable form the dividing line between “not storing” and “retaining memories.”
References
- “Rabbit’s new AI agent doesn’t need an R1 to run”, by Emma Roth — The Verge, 2026-09-22T20:52:40.000Z (ARR)
- Source URL: https://www.theverge.com/ai-artificial-intelligence/999094/rabbit-ai-agent-os3
Frequently Asked Questions
- Can Rabbit OS3 be used without owning R1
- Yes. OS3 runs in the cloud while executing operations locally on existing Windows, Mac, and Linux devices. It can be accessed via a dedicated site and Telegram and iMessage, and R1 is not required. Up to five devices can be registered per account.
- How does OS3 execute tasks
- OS3 automatically identifies the necessary elements from registered devices, files, applications, and AI models and carries out the work. Users do not need to specify the environment individually, and the design autonomously completes a series of tasks across multiple materials and apps.
- How does OS3 handle data
- Rabbit says it does not store, duplicate, use, or sell data, while also stating that chat content and memories derived from it are retained on its servers. Linked external AI providers handle data under their own privacy policies, so users need to check each policy.
Comments