1
1
Jesse Lyu, the visionary behind the distinctive orange-red Rabbit R1, vehemently refutes the widespread perception that his much-hyped AI gadget was a commercial or technological failure. Launched with considerable fanfare at CES 2024, the Rabbit R1 quickly became a symbol of the nascent "gold rush" in dedicated AI hardware, aiming to establish itself as a personal virtual assistant capable of executing complex tasks like booking an Uber or ordering food via DoorDash simply through voice commands. However, its initial reception was far from glowing, with many prominent tech publications, including WIRED, delivering scathing reviews that highlighted the device’s unreliability and the limited functionality of its core agentic features, famously earning a 3/10 score from WIRED.
Despite the critical backlash, Lyu maintains that the Rabbit R1 achieved significant monetary success. He asserts that the company successfully shipped over 100,000 units of the Teenage Engineering-designed device worldwide, boasting an impressive return rate of less than 5 percent. Furthermore, Lyu claims the hardware yielded a substantial profit margin of 45 to 50 percent, a notable achievement considering Rabbit opted against a recurring subscription model for its AI features, a common revenue strategy among many of its competitors in the burgeoning AI space.
Now, with the capabilities of AI assistants having advanced considerably since the R1’s debut, and the concept of intelligent agents becoming a focal point of innovation, Rabbit is once again making headlines. The company recently unveiled OS3, a new standalone "agentic operating system" designed for expansive multi-screen compatibility. While OS3 will arrive as a software update for existing R1 devices, its core innovation lies in its independence; users will not require an R1 to access its features. Instead, OS3 can be seamlessly integrated and utilized across a desktop browser, within the Telegram messaging app on a smartphone, or even directly through iMessage.

This move represents a striking and strategic pivot for Rabbit. The company initially championed the superiority of dedicated handheld hardware over the perceived limitations of smartphone applications for AI interaction. However, the current technological landscape has significantly matured, creating a "ripened" environment for sophisticated agentic software. Indeed, the concept of AI agents capable of autonomously handling tasks is now a dominant topic of discussion in Silicon Valley. Lyu subtly nods to Rabbit’s early experiences by referencing contemporary challenges faced by larger tech entities. He recalls the recent news of Meta’s new Muse agent being blocked from crawling Amazon to facilitate user shopping via a virtual machine. With a knowing smirk, Lyu points out, "We’ve been through all of this a year and a half ago," alluding to the R1’s own controversial use of a virtual machine to execute app actions on behalf of its users. This suggests that Rabbit, perhaps, was simply ahead of its time, encountering hurdles that larger players are only now beginning to navigate.
Delving into the company’s operational realities, Lyu revealed that Rabbit has raised approximately $60 million in total funding and maintains a remarkably lean team of just 15 employees. He acknowledges that this limited resource pool, both in terms of personnel and capital, presents a significant challenge when competing against "other AI studios that raised billions of dollars." The path to this current announcement was not without its trials. Around a year ago, Rabbit faced severe financial strain when a planned launch in India was unexpectedly thwarted by regulatory obstacles. This period of difficulty led to a handful of employees going on strike, alleging unpaid wages. Lyu expresses his respect for their decision, confirming that the company has since secured additional funding and is now on "stronger footing," asserting that "there are no issues with the company’s health."
In the wake of the R1’s initial lukewarm reception, the Rabbit team made a deliberate decision to prioritize software development over rushing new hardware to market. This commitment resulted in the deployment of approximately 50 software updates for the R1 over the past year and a half, significantly refining the user experience. Lyu highlights the integral role of Rabbit’s dedicated Discord community, comprising over 12,000 active users, who provide invaluable feedback that directly influences new feature implementations. This community-driven approach has fostered an environment of openness, exemplified by the R1’s ability to integrate third-party AI agents such as OpenClaw or Hermes AI, allowing users to move beyond Rabbit’s proprietary models. Lyu proudly states, "Today, if you look at our R1 device, it’s completely different from day one… We’re more open; the device is yours, you decide what to do with the device." He even demonstrated recent R1 orders on his phone during the interview, indicating that the device continues to sell through Rabbit’s website and major retailers like Amazon, Target, and Best Buy, albeit with dwindling stock.
Looking ahead, Rabbit is ceasing the manufacturing of the R1. There are currently no plans for an R2. Instead, the company’s next hardware release, anticipated "within months," will be what Lyu describes as a "Cyberdeck" – a dedicated "vibe coding" machine specifically designed to run the new OS3 platform. This suggests a more specialized, creator-focused hardware offering that aligns with the software-centric vision of OS3.

So, what exactly is OS3? Lyu positions it as an open, cross-platform agentic system engineered to operate fluidly across the existing devices users already own. The platform resides primarily in the cloud but establishes direct connections to desktop software, local files, and web pages through a localized agent node that can be installed on Windows, Mac, or Linux operating systems. At its core, OS3 leverages Rabbit’s proprietary DLAM (Direct Large Action Model) technology, enabling it to execute local software tasks and even autonomously write and debug code. A single user account can link up to five machines, with OS3 intelligently determining which specific device is best suited to execute a given task behind the scenes.
The interface of OS3 is designed for simplicity and efficiency: a single, continuous chat stream, eschewing disjointed sidebars often found in conventional chat applications. Lyu frequently refers to it as a "brain dump" interface, where users can simply articulate their intent, and the system intelligently coordinates the necessary AI models and tools to achieve the desired outcome. A key feature is its "universal skill compatibility," allowing users to paste a public URL for any third-party agent skill directly into the chat, which OS3 can then automatically install without requiring any technical expertise.
Rabbit’s approach to powering OS3 is a "Bring Your Own Key" (BYOK) model. This means there are no monthly subscription fees from Rabbit itself. Instead, users provide their own API keys from preferred model providers like OpenAI or Anthropic, or even link local open-source models, offering significant flexibility and control over costs and model choice.
Lyu illustrates OS3’s practical utility with personal anecdotes. He describes how he typically receives weekly Excel reports detailing R1 sales from vendors. While he might initially check these on his phone, he previously had to switch to his PC to open the document and manually input the data into a master sheet. With OS3, the Rabbit agent deployed on his PC can receive a command via Telegram on his phone to handle this routine data entry, automating the entire process. In another instance, Lyu, facing a full day of video shooting and limited Slack access, instructed OS3 to represent him on the platform. The agent, transparently identifying itself as Lyu’s agent, engaged in a 50-minute conversation with an engineer, even accessing files on Lyu’s computer to provide answers to technical questions.

Regarding data privacy, Rabbit assures users that the local agent does not store, copy, or sell any user data. While conversations and memory logs are maintained on Rabbit’s servers, the actual content of user files remains securely on their local machines. However, a crucial caveat is that when a task necessitates model reasoning, the relevant prompts are routed through Rabbit’s servers to the user’s chosen model provider, where that data is then processed under the provider’s specific terms and conditions.
The launch of OS3, accessible to anyone starting today, subtly but powerfully refutes the initial premise that propelled Rabbit into the spotlight in 2024. It suggests a fundamental realization: the market might not have needed another dedicated gadget in a pocket to interact with AI. Instead, users already possess the most versatile tools for the job – their existing devices. Rabbit’s evolution from a hardware-centric startup to a software-first agentic platform reflects a deeper understanding of user needs and the rapidly evolving landscape of artificial intelligence.