Jesse Lyu believes the Rabbit R1 was not a failure.
The device arrived among the first wave of dedicated AI hardware intended to function as a virtual assistant. Users could speak into the R1 to book rides or order food. Following a launch at CES 2024, reviews were scathing. WIRED gave it a 3/10 score because the agent features performed poorly.
AI assistant capabilities have improved vastly since then. Rabbit is promoting its return.
On Tuesday, the company announced OS3. This standalone agentic operating system runs across multiple screens. It updates the existing R1 hardware, but you do not need the gadget to use it. Access is possible through a desktop browser, Telegram on a phone, or iMessage.
This is a sharp pivot for a firm that once championed handheld hardware over smartphone apps. The environment for this software has matured.
Agents that handle tasks are the main topic in Silicon Valley. Meta recently banned its new Muse agent from crawling Amazon to shop on a user’s behalf. Lyu mentions this with a smirk. He notes, “We’ve been through all of this a year and a half ago,” referencing the R1’s controversial use of a virtual machine to execute app actions. Rabbit was perhaps too far ahead of the curve.
Lyu states the Rabbit R1 was financially successful. The company does not offer a subscription for its AI features, unlike many rivals. He claims they delivered more than 100,000 orange-red gadgets worldwide with return rates below 5 percent. The hardware margin was roughly 45 to 50 percent.
“Internally, from an engineering perspective—we didn’t make the wrong bet,” Lyu says. “I think the challenge is really being extremely limited in resources, team-wise and money-wise, compared to other AI studios that raised billions of dollars.” Rabbit has raised around $60 million in total and currently employs 15 people.
After poor reviews at launch, the company made tough choices. Rabbit avoided expanding the team to stay lean, but conditions were dire. A planned launch in India faced regulatory hurdles, creating financial strain. A handful of employees went on strike over unpaid wages. Lyu respected their decision. The firm secured more funding and is on stronger footing. “There are no issues with the company’s health,” he says.
The team avoided rushing new hardware. Instead, they focused on software updates to improve the R1 experience. Lyu says Rabbit has deployed around 50 updates in the past year and a half. A dedicated Discord community of more than 12,000 users provides feedback. Users can now integrate OpenClaw or Hermes AI agents into the R1 instead of using Rabbit’s proprietary models.
“Today, if you look at our R1 device, it’s completely different from day one,” he says. “We’re more open; the device is yours, you decide what to do with the device.”
People are still buying the R1. Lyu showed his phone screen to a webcam feed displaying four or five orders from that morning sold through Rabbit’s website. He notes this excludes Amazon, Target, and Best Buy sales.
For now. Rabbit has stopped manufacturing the R1. Existing stock is dwindling. The company focuses on the OS3 software platform, which works on hardware you already own. If you have an R1, you can use it as an input and output gadget for OS3. Lyu says Rabbit has no plans for an R2. The next hardware release, arriving within months, is a Cyberdeck. This dedicated “vibe coding” machine will run OS3.
OS3 is an open, cross-platform agentic system that operates on your existing devices. The platform lives in the cloud but connects directly to desktop software, files, and web pages through a local agent node installed on Windows, Mac, or Linux. Using Rabbit’s DLAM (Direct Large Action Model) technology, it executes local software tasks and autonomously writes and debugs code. One account links up to five machines, with OS3 deciding behind the scenes which device should execute a given task.
The interface is a single, continuous chat stream. There is no sidebar of disjointed chat threads. Lyu calls it a “brain dump” interface. Paste your intent, and the system coordinates the models and tools needed to complete it. It supports universal skill compatibility. Paste a public URL for any third-party agent skill into the chat, and OS3 can install it automatically without users needing technical know-how.
Rabbit uses a “Bring Your Own Key” (BYOK) approach to power OS3. There is no monthly subscription fee, but users supply their own API keys from providers like OpenAI or Anthropic, or link local open source models.
Lyu has used OS3 to handle various tasks. Vendors send him an Excel report for R1 sales every week. He checks it on his phone but returns to his PC to open the document and manually add the data into a master sheet. With OS3, since the Rabbit agent is deployed on his PC, he can send a message through Telegram on his phone to handle the task. The system completes the routine action.
Recently, Lyu was heading to a video shoot that would occupy most of his day. He was not reachable on Slack. He asked OS3 to represent him. The agent had a 50-minute conversation with an engineer. It announced itself as Lyu’s agent, so the engineer was aware. The agent could even access files on Lyu’s computer to help answer questions.
Rabbit states the local agent does not store, copy, or sell data. While conversations and memory logs reside on Rabbit’s servers, file contents stay on your machine. The caveat is that when a task requires model reasoning, the relevant prompts pass through Rabbit’s servers to your model provider, which then processes that data under its own terms.
Anyone can try OS3 starting today. This quietly refutes the premise that launched Rabbit into the spotlight in 2024. People never really needed another gadget in their pocket to chat with AI. They already have the best tool for the job.
What it means
For people making things, the shift is from owning a specific hardware device to owning a flexible software layer. The R1 is no longer the primary tool; it is now just one input method among many. Users can direct a single agent to work on their PC, phone, or browser based on where the data lives. This removes the friction of managing separate apps for different platforms. You simply tell the agent what to do, and it figures out which machine to use. This approach allows creators to integrate third-party tools directly without building custom integrations for each one.
