Rho-1: Reka combines text, images, video and robotic actions in a single model with 19 billion parameters

Rather than an orchestrator that delegates tasks to specialised models, Reka presents a single network that comprehends, generates and acts, with a distilled version capable of rendering 5 seconds of video in approximately one second.

Reka publishes a research overview of Rho-1, an "omni-reasoning" model of 19 billion parameters trained from scratch. His idea: to replace agent chains, where a central model assigns each task to a specialist (image, video, detection), with a single network in which text, visual content and robotic commands are all tokens of the same context. An understanding branch and a diffusion generation branch share attention and the same cache, which serves as persistent state memory from one round of conversation to the next.

In the demonstration, the model draws a lighthouse, frames it, animates it into a video, changes the weather in the sequence and then explains the difference between the two versions, without calling a tool or second model. Reka announces a video generation at 0.79 times the real time for the basic version, and a distilled variant that reduces the denoising from 99 to 8 steps and produces a 5.3-second clip in just over a second. The company also presents Rho-1 as a real-time controllable world model and as a basis for controlling robots, which it sees as a prerequisite for a general "physical AI".