Reka AI Labs says it is releasing a research preview of Rho-1, which the company describes as a 19B omni model that understands and generates text, images, video and robot actions inside a single neural network.
The company's pitch is less about another model launch than about collapsing several AI jobs into one system. In its launch posts, Reka says most AI systems are agentic pipelines that hand work to specialist models, while Rho-1 is meant to handle text, images and video in one conversation from what it calls a shared world state. Reka staffer Marco Atzeni used similar framing, writing that Rho-1 works without an agentic loop or tool calls.
What Reka says Rho-1 can do
Reka also describes Rho-1 as a steerable world model that can generate continuous video in real time and change course mid-generation when new instructions arrive.
In a separate post, the company says Rho-1 "closes the loop for physical AI". Reka's claim is that the same model weights predicting what a camera should see next can also dictate joint actuation, rather than relying on an external planner querying a separate world model.
How Reka is framing the preview
Reka says Rho-1 was trained from scratch on 320 H100 GPUs in about three months, which it calls a small fraction of the compute behind today's frontier models.
Staff posts add some caveats to the launch. Dani Yogatama, a Reka staffer, wrote that the team is "very early on the compute curve" and "candid about current limits". Another staffer, Xabier Artetxe, said the company was sharing unedited screen recordings of Rho-1 "exactly as it ran", adding that there were no specialist models operating behind the scenes.
On its product page, Reka describes Rho-1 as "an omni-reasoning model that understands, simulates, and acts". For now, though, the release is explicitly a research preview, and the company's own staff posts frame it as an early look rather than a finished system ready to speak for itself.
Claim support
Comments