Reka AI released Rho-1 on October 5, 2026, saying it processes and generates text, images, video, and robot control actions in a single neural network. It is the company's first multimodal update since shipping Reka Core in April 2024.
Reka AI reported a 19-billion-parameter model, measured on benchmark scores, which compares with its earlier multimodal language model, Reka Core. That model competed with GPT-4, Claude 3, and Gemini Ultra on benchmarks.
Rho-1 is built on an inverse dynamics model and targets robot control, with availability beginning in the research preview phase, initially for developers and researchers.
"The same weights that predict camera images also drive robot movements," said Matthias Bastian, a researcher at Reka AI. The model generates continuous video in real time and responds to new instructions on the fly without restarting.
The announcement follows a broader push in AI research toward so-called world models, as described by Reka AI. The company did not say how it will scale the model further, and it raised the question of how to handle scarce robot training data. The model trained on 320 H100 GPUs over about three months.
Source: thedecoder
