Reka’s Rho-1 combines video, images, text and robot actions

Reka describes Rho-1 as a 19-billion-parameter model trained from scratch.[1] The company says it can understand and generate text, images and video, reason over them, and take actions within one neural network.[1] Text, vision and robotic actions are represented as tokens in a shared context window, according to Reka.[1] MarkTechPost reports that the streams share attention and use the same key-value cache.[2] Reka says new instructions can arrive at any moment.[1]

In one demonstration, Rho-1 generated a scene, located an object, animated and edited the video, then explained the changes in five turns.[1] Reka says this used one model and one conversation state, without a tool call or second model.[1] The robotics demonstration was narrower: MarkTechPost reports that Rho-1 produced seven action channels in a LIBERO simulation episode.[2] The Decoder reports Reka’s claim that the same weights predict camera images and drive robot movements.[3] RuntimeWire noted that the announcement did not demonstrate Rho-1 controlling a physical robot.[4] The simulation result therefore should not be read as a physical-robot deployment demonstration.[2][4]

Reka reports that the base model generates video at a median 0.79 times real time and begins a watchable stream in roughly six seconds.[1] The company also says it returned a 5.3-second video clip in about a second.[1] A distilled version reduces video denoising from 99 steps to eight, according to Reka.[1] Those figures describe different reported measures; they do not establish that every task or output runs at the same speed.[1] Rho-1’s native video output is currently limited to 672 by 384 pixels.[1] Teams assessing it for video workflows should weigh these reported speeds against resolution needs and task-specific results.[1]

Reka says Rho-1 can drift structurally during long video rollouts, even while retaining visual detail.[1] The company says object detection and coordinate grounding work reliably on static images but not yet across video.[1] It also describes targeted visual editing as nascent and consistency across diverse prompts as brittle.[1] The research preview has no public weights, API or pricing yet, according to Reka’s launch materials.[2] Reka reports that the model was trained on 320 H100 GPUs over about three months.[3] For now, the announcement offers evidence of a unified research direction, not enough information to plan a public API integration.[3][2]