Reka releases Rho-1, a 19B omni-reasoning model unifying text, images, video and robot actions

On October 5, San Francisco AI lab Reka AI released a research preview of Rho-1, a 19B-parameter omni-reasoning model. Unlike mainstream agentic pipelines that route work to modality-specific models, Rho-1 unifies text, vision and robotic actions as tokens in a single context window, with one set of weights predicting the next camera frame and outputting joint actions.

In a demo, Rho-1 first draws a seaside lighthouse, then draws a bounding box around it, generates a drone-push-in video, edits the scene to a blizzard, and finally answers questions about the differences between two clips in text. Every turn reads and writes the same shared state, with no second model call and no tool calls in between. The base Rho-1 generates video at 0.79x real-time median, and a distilled variant cuts the denoising trajectory from 99 to 8 steps, producing a 5.3-second clip in about a second.

To bypass scarce robot data, Reka trained an inverse dynamics model that extracts control signals from ordinary internet video, enabling web-scale video as action-labeled pretraining data. Rho-1 uses a symmetric architecture where each transformer block splits compute into an understanding stream (language and visual parsing) and a generation stream (denoising latents into images and video); the two streams do not run in isolation.

Training used only 320 H100 GPUs over about three months, a small fraction of the compute behind modern frontier language and video models. Native video resolution is capped at 672×384. The team is candid about limitations including long-horizon structural drift, cross-video object grounding instability and brittle targeted editing. Rho-1 is currently available as a research preview, with the team engaging enterprise partners on embodied robotics, interactive simulation and closed-loop vision-action systems.