Reka previews Rho-1, a 19B single-network model unifying text, image, video generation and robotic actions

What changed

Reka AI released a research preview of Rho-1 on October 5. Rho-1 is a 19B omni-reasoning model trained from scratch that represents text, visual content and robotic actions as tokens in one context and uses a single network for understanding, text/image/video generation, continuous world simulation and action output rather than routing tasks to separate specialist models. Reka says the checkpoint was trained for about three months on 320 H100 GPUs, and a distilled variant reduces denoising from 99 steps to 8. The company also calls it a proof of concept and lists long-horizon structural drift, weak temporal grounding, brittle editing and 672×384 native video resolution as current limitations.