Reka AI has introduced a research preview of Rho-1, a 19-billion-parameter omni-model capable of natively handling text, image, continuous real-time video, and robotic control actions. Unlike traditional modular architectures that route specific tasks to external tools, Rho-1 executes all input and output modalities as tokens within a single shared context window.
To overcome the shortage of physical robotics data, the company developed an inverse dynamics model designed to extract control signals directly from standard internet videos. The model was trained over approximately three months using a cluster of 320 Nvidia H100 GPUs.
This release advances research into integrated world models capable of direct physical interaction. By enabling the same neural weights that process visual inputs to directly generate motor commands, Reka AI aims to simplify embodied AI architectures and reduce instruction-processing latency.
Why it matters
Single-model native multi-modality removes the need for complex tool-calling pipelines in robotics.
Extracting robot control signals from web video offers a path around physical data bottlenecks.
19B model size demonstrates efficient multimodal capabilities without massive compute overhead.
Source: the-decoder.com



