Rho-1 is Reka AI's newly released research-preview model. On October 5, 2026, The Decoder reported on this 19-billion-parameter "omni" model: text, images, video, and robot control actions are processed and generated inside a single neural network, with no external tool calls and no routing of tasks to separate specialist models. Every modality is placed into one shared context window and computed as a single stream of tokens.
One network, four jobs — the hard part is data, not architecture
According to the report, Rho-1 can generate continuous video in real time, accept new instructions mid-run without restarting, and use the same weights that predict camera images to directly output robot movements. Scarce robot data is the classic obstacle for such models; Reka's answer is to first train an inverse dynamics model that infers control signals from ordinary internet videos, turning huge amounts of unlabeled footage into usable training material. The model itself was trained on 320 H100 GPUs over roughly three months. In Reka's official roadmap this is no detour: after merging talent with Moonvalley, the company has been pursuing "omni world models" whose single architecture simulates, reasons, and predicts actions — and Rho-1 is the first visible instance of that line.
How it differs from today's mainstream multimodal approach
Mainstream vision-language models are usually "many inputs, one output": they look at images and video and answer in text, while image generation or robot control is bolted on through other models, with latency and information loss at every seam. Rho-1 bets on folding understanding and generation into one shared representation, letting the model internally rehearse how the scene will change before deciding how to act. For robotics, planning and execution stop being a relay between two systems; for interactive media, video can respond to new instructions while it is being generated. The RobCo robotics funding story we covered shows capital chasing robot bodies; work like Rho-1 supplies the other half — a more general shared "brain" for those bodies.
Too early to grade it
Public information about Rho-1 currently comes mainly from one outlet's report and demo video: no published benchmark data, no weights, no trial access, and at 19 billion parameters it is plainly not aiming at the frontier scoreboard. Its value is closer to a route validation — can a single network carry understanding, generation, and action at once, and can the inverse-dynamics data route scale? Wait for Reka's technical report or a reproducible demo before judging how far it is from a usable omni model.
