Reka AI released a research preview of Rho-1, a 19-billion-parameter omni-model capable of processing and generating text, images, video, and robot control actions within a single neural network. Unlike conventional AI systems that route different tasks to specialized models, Rho-1 manages all modalities as tokens inside one shared context window, operating without external tool calls or separate models. The system is designed to generate continuous video in real time while responding dynamically to new instructions without needing a system restart.
To bridge the gap caused by scarce robotic training data, Reka AI developed an inverse dynamics model that extracts control signals directly from ordinary internet videos. The model utilized the same underlying weights to predict camera imagery and drive robotic movements, having completed its training on 320 H100 GPUs over a period of approximately three months. This release follows Reka AI's previous launch of Reka Core in April 2024, a multimodal language model that competed with industry benchmarks like GPT-4, Claude 3, and Gemini Ultra, aligning with a broader push in artificial intelligence research toward unified world models.
This development matters because it demonstrates the feasibility of combining multiple distinct modalities and robot control into a single architecture, reducing reliance on specialized routing models and scarce training datasets.