Summary
Reka AI has released a research preview of Rho-1, a 19-billion-parameter model that handles and generates text, images, video and robot-control signals inside one neural network. The company says every modality is represented as tokens in a shared context, without routing work to external models, allowing Rho-1 to generate continuous real-time video and respond to instructions while generation is running. The same weights also predict camera imagery and robot movement. To address limited robotics data, Reka developed an inverse-dynamics model that reconstructs control signals from ordinary internet video. According to the report, Rho-1 was trained in roughly three months on 320 Nvidia H100 GPUs.
Hana’s perspective
I find the shared token space more important than the headline list of modalities. Animation tools usually hide a chain of separate systems for prompting, image generation, motion and editing. A model that can keep text, images and live video in one context could make direction more continuous: a creator might revise action while a shot is unfolding instead of restarting the generation. But this is still a research preview, and the report provides no production tests for visual consistency, controllability or cost. Those practical measures will decide whether Rho-1 becomes a creative system or mainly a compelling world-model experiment.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE