Reka Rho-1 Unifies Video and Robot Actions in Preview

Reka’s October 5, 2026, announcement of Rho-1 introduced a 19-billion-parameter research preview. The model understands and generates text, images and video, reasons across them, and produces robot actions within one neural network. Its central idea is to retain visual and action history in a shared context as a task unfolds.

The RuntimeWire account of the October 5 preview identifies the robot-action example as a LIBERO simulation, rather than a physical robot trial. Researchers can examine the published demonstrations, while prospective collaborators are invited to contact Reka. The launch offers no public model download, self-service API or pricing for Rho-1.

One context carries the scene between tasks

Rho-1’s clearest demonstration begins with a generated image of a red-and-white lighthouse on a rocky headland. In the same conversation, the model marks the lighthouse, animates an approach toward it, changes the resulting video into a snowstorm and answers a question about the edit. The sequence shows image generation, localization, video creation, editing and a text response connected through the accumulated model state.

In a conventional agent pipeline, a planning model sends selected instructions and context to specialist systems for tasks such as detection or video generation. Rho-1 instead represents those operations within one network and context window. Text and high-level commands use discrete tokens; image and video representations, proprioception and robot actions use continuous tokens. That lets an output become part of the state available to later steps in the session.

The network still divides work internally. An understanding stream handles language and visual parsing, while a generation stream renders images and video; both use shared attention and a common cache of preceding context. When the session moves from an image to video, the generation stream can draw on the image representation already in context. The demonstrated benefit is continuity across the sequence, although the launch does not establish how consistently it holds across varied prompts and longer sessions.

Robot actions share the visual prediction

The robotics demonstration applies that architecture to a manipulation task. In a LIBERO simulation episode, Rho-1 predicts a robot’s wrist-camera view and emits seven action channels from the same model state. The point is the relationship between a predicted observation and a proposed movement: both are produced within the network’s shared context, rather than passed between a separate world model and control policy.

An inverse-dynamics model offers another route for preparing training data. It infers likely control signals from raw video, and Rho-1 can ingest those signals as action tokens. This could make footage without recorded robot commands useful for action-related training. An inferred signal remains an estimate of what caused a movement, however; the published simulation does not show that the resulting actions are reliable on physical hardware.

Claims and the evidence behind them

The release combines an architectural description, recorded demonstrations and company-run measurements. Each supports a different conclusion about the preview:

Modalities — demonstrated sequence: The lighthouse session connects image creation, object localization, video generation, an edit and a text answer. It shows these tasks sharing context in one recorded conversation, without measuring consistency across an independent prompt set.
Model size — disclosed specification: Rho-1 has 19 billion parameters and was trained from scratch. Parameter count describes the model’s scale; it does not establish the quality of its images, videos or actions.
Robot actions — simulation evidence: The LIBERO episode displays an observed wrist view, a predicted view and seven emitted action channels. It demonstrates action output in simulation, with no physical robot trial in the launch.
Latency — internal measurement: The published timings cover a base model and a distilled variant. They are useful descriptions of the company’s tests, but independent, task-matched replication is not provided.
Quality — limited comparison: Reduced denoising with little reported quality loss is a company assessment. The demonstrations do not supply an independent benchmark against specialist image, video and robotics systems under common conditions.
Access — research preview: Rho-1 is presented to prospective collaborators through a contact invitation. Published examples and a preview designation do not provide developers with an openly downloadable model or a self-service endpoint.

Speed and the boundary of the preview

The MarkTechPost breakdown of the company’s tests gives the base model’s median video generation speed as 0.79 times real time and says a distilled variant reduced denoising from 99 steps to eight, producing a 5.3-second clip in about one second. These figures describe internal measurements. The displayed first-clip race likewise places a measured Rho-1 time beside an illustrative agent-pipeline time, rather than results from two systems tested under matched conditions.

The preview also has limits tied directly to its shared visual state. Extended video can drift into an incompatible scene layout; object grounding that works on static images remains unreliable across video; and targeted edits can be brittle across prompts. Native video output is capped at 672 by 384 pixels. Those weaknesses matter most when a later instruction or robot action depends on a scene remaining stable over time.

For developers, access terms and independent evaluations are the next consequential pieces of evidence. Tests that apply the same visual tasks to Rho-1 and specialist pipelines, and that assess action output on physical hardware, would show how far the demonstrated shared context carries beyond this preview.