Reka unveils rho-1: a unified model for text, video, and robot actions

Reka has introduced Rho-1, a 19-billion-parameter omni-reasoning model capable of understanding, generating, and executing tasks across text, images, video, and robot actions. Unlike traditional multimodal systems that rely on multiple specialized models, Rho-1 operates as a single neural network, eliminating the need for handoffs between different models. The model can generate a lighthouse, box it, animate it, and explain the changes in just five turns without external tools. Rho-1 uses two native formats for input and output, with transformer blocks containing two expert weight streams—one for understanding and one for generation. The model generates video at 0.79x real-time, with a first clip appearing in about 7 seconds, outperforming a multi-agent pipeline that took 13.8 seconds. A distilled version of the model reduces denoising steps from 99 to 8, achieving a 5.3-second clip in under a second. The model also supports continuous streaming and mid-rollout updates, demonstrating its versatility in robotics and simulation tasks.