Reka Releases Rho-1 to Combine Video Creation and Robot Actions in One Model

Creating a scene, turning it into video and asking what changed can now happen inside one model in Reka’s research preview. Released October 5, 2026, Rho-1 is a 19-billion-parameter system that the company says combines text, images, video and robot actions in a shared context, rather than passing work between separate models.

The lighthouse stays inside the conversation

Reka’s five-turn demonstration starts with a red-and-white lighthouse on a rocky coast. The user asks the model to mark the lighthouse, animate a drone’s approach, replace the weather with a snowstorm and explain the difference between the videos. Each response uses the same model and accumulated conversation state.

Underneath that workflow are two connected streams: one handles understanding and text responses, while the other generates visual material. Both use shared attention and a common KV cache—the stored internal information carried forward through the session. Reka says there is no tool call or second model producing those responses.

The distinction is more than keeping a chat history. When Rho-1 animates the lighthouse, it uses the image representation already in context rather than re-encoding an exported picture. When it explains an edit, it reads the internal representation that produced the change, rather than sending finished frames through a separate captioning model.

Object marking takes a different path through that architecture. The model outputs coordinates for the box using the scene already in its context, without a separate object detector. Reka also says the model itself plans the shot before rendering it, rather than relying on an added prompt-enhancement system.

Flash reduces the work between prompt and pixels

Reka reports that base Rho-1 generates video at a median 0.79 times real time, with a watchable stream beginning after roughly six seconds. A faster, distilled variant called Rho-1 Flash reduces denoising—the repeated steps that turn a noisy representation into an image or clip—with what Reka describes as minimal quality loss.

In the company’s shown Flash examples, 5.3-second clips render in 1.11–2.01 seconds as prompts change a dog’s breed, clothing and surroundings. These are company-reported timings. Reka’s broader speed comparisons come from internal testing; its pipeline comparison is illustrative, and the lighthouse playback shortens generation waits.

Instructions can arrive while the world is moving

Rho-1 also supports continuous generated video that can take new instructions without restarting, according to Reka. One demonstration branches an aerial scene into “bank left” and “bank right” continuations. Both share the opening half-second before following different paths. Another shows a simulated robot arm picking up and putting down a ball as commands arrive.

During a live rollout, a new command enters the understanding stream and updates the stored state. The generation stream then renders the changed trajectory. Reka proposes using such reactive environments to test autonomous-driving systems or train robot-control policies, with developers changing conditions while a simulation runs.

Robot control extends the shared-model idea beyond generating footage. Reka says the same model weights that predict camera images also drive robot movements. To address scarce robot training data, the company built an inverse dynamics model that derives control signals from ordinary internet videos, The Decoder reports.

Reka trained Rho-1 from scratch. The training run used 320 H100 GPUs over about three months, according to The Decoder, covering the model that combines media generation and understanding with action output.