The Pulse
Reka Puts Text, Video and Robot Control in 19B Rho-1
Reka introduced Rho-1, a 19-billion-parameter research-preview model that processes and generates text, images, video and robot actions in one network. The company reports real-time video generation and simulated robot-control demonstration
AI.info Team · Written with AI ·
“Rho-1 is a functional proof-of-concept and an architectural direction, not a finished product.”
Reka AI
Reka introduced Rho-1 on October 5 as a 19-billion-parameter model that handles text, images, video and robot actions within one neural network. The company says it was trained from scratch and combines tasks that are often split between a language model, an image or video generator, and a separate robotics policy. Reka is releasing it as a research preview, not a finished product.
One model shares a running state
Rho-1 treats text, visual information and robotic actions as tokens in a shared context. Reka says the same model can draw an image, locate an object in it, animate it, edit the resulting video and answer questions about the change without handing each step to a separate model. Its post describes the model as maintaining one state across those turns, rather than starting over for each task.
The architecture has two streams within each transformer block: an understanding stream for language and visual parsing, and a generation stream for producing images and video. Both use shared attention and a common cache, according to Reka. Training combines next-token prediction for discrete sequences with flow matching for continuous outputs such as image and video latents and robot actions.
Video speed comes with a resolution limit
Reka reports that its base model generates video at a median 0.79 times real time, with a watchable stream starting in about six seconds. A distilled version reduces the denoising process from 99 steps to eight; the company says it produced a 5.3-second clip in about one second in internal testing. Those figures are Reka’s own measurements, not results from an independent evaluation.
The release also spells out trade-offs. Reka says long video rollouts can drift in structure even while their images remain detailed, and that object grounding across video is not yet reliable. Targeted edits can be brittle, and native video output is capped at 672 by 384 pixels. Those constraints place the launch closer to an architecture demonstration than a ready-to-deploy video system.
Robot control remains a simulation demonstration
Rho-1’s robotics examples show the model predicting future camera views while emitting actions for simulated tasks. Reka says the model’s shared representation lets it use predicted future scenes to inform the movements it outputs. The post includes a LIBERO simulation episode and generated examples of objects being picked up, but does not present evidence of deployment on a physical robot.
To address the shortage of robot-control logs, Reka says it pairs Rho-1 with an inverse dynamics model that infers control signals from ordinary video. The aim is to turn video without recorded joint movements or steering inputs into training examples that include actions. The Rho-1 post does not report a robotics success rate or a standardized comparison for that approach.
Reka is asking builders to collaborate
The company says the checkpoint was trained on 320 H100 GPUs over three months. Reka describes the result as a proof of concept and invites teams working on robotics, interactive simulations and vision-action systems to contact it about collaboration. The release does not announce open model weights or general public access.
For now, Rho-1’s most concrete claims are about what one model can do in demonstrations—and what still fails: video grounding, stable edits, long-horizon consistency and resolution beyond 672 by 384 pixels.