The Decoder reported on 5 October that Reka AI has opened a research preview of Rho-1, a 19-billion-parameter model. One network handles text, images, video and robot control actions, the outlet said. It turns all of them into small data units called tokens and keeps them in a shared working space, without calling tools or other models. The Decoder also said the model makes video live and accepts new instructions mid-run. Training took about three months on 320 H100 chips.
The Decoder said Reka built an inverse dynamics model because robot training data is scarce. Reka's research page describes one, called RIDM. Given two video frames, it infers which move happened, like a caption for motion. Game footage with engine-supplied labels trained it. One version first converts frames to optical flow, a map of how pixels shift, using RAFT-small, and Reka says this lets it work on real video. Reka lists failures, such as false turns on dark frames. The page text does not state how RIDM relates to Rho-1. The sources give no Rho-1 benchmarks or access terms.
The Decoder frames Rho-1 as one model doing what most systems split across several; no benchmarks or outside checks back that yet. Buyers can ask for proof it works and for how it fails. Reka's Apache 2.0 licence covers only RIDM's weights and code, not Rho-1, and the RAFT weights its flow version needs may limit commercial use.