Reka's 19B Rho-1 folds video generation and robot actions into one shared context

The shared state is the strongest part of the engineering claim. In the familiar design, an agent links a chain of specialist models ^[6]. In Rho-1, Reka says, each turn reads from and writes to one shared state, with no second model or tool call in the sequence ^[7]. The demo conversation starts with a generated lighthouse image. It then locates an object in the scene, animates it, changes the weather in the video and answers a question about the edit ^[7]. Carrying one scene through all of those turns is clean design, and it is the part of the October 5th preview I would test first ^[1].

The robotics case reuses the same weights. The model predicts future camera frames and emits joint actions ^[3]. Reka's bet is that a system that predicts what it will see and produces the action can cut the handoffs between a world model, a planner and a robot policy ^[5]. Robot training data is scarce, so Reka built an inverse dynamics model that pulls control signals out of ordinary internet video, according to The Decoder ^[8]. Those action labels are inferred from footage. I'd want to see how far the labeller's errors carry into the joint actions the model emits. The announcement does not show Rho-1 controlling a physical robot ^[3]. Simulation is the right place to start. Nothing in LIBERO breaks when it falls over.

The speed figures come from Reka's own tests ^[13]. The base model generates video at a median 0.79 times real time, and a stream starts in roughly six seconds ^[9]. The Decoder describes the output as continuous real-time video that takes new instructions without restarting ^[19]. A median tells you about the middle run, and a live stream stalls on the slow ones. The distilled variant cuts the denoising path from 99 steps to eight and produced a 5.3-second clip in about one second ^[10]. That is about 12 times fewer steps ^[17], and output at roughly five seconds of video per second of compute ^[18]. For either figure to transfer, a buyer's prompts, clip lengths and serving hardware would have to match Reka's test setup.

Reka published its failure modes with the preview, and that is good practice. Long video rollouts drift in structure, and targeted edits remain brittle across prompts ^[11]. The company attributes its limitations largely to training scale and data, and expects both to improve as it scales them ^[20]. RuntimeWire noted that the preview has not established that scaling will improve them ^[21].