Reka releases 19B Rho-1 research preview for unified processing of text and images, video, and robot actions

AT A GLANCE

Reka releases a 19000000000 parameter model Rho-1 focused on text and images, video generation, and robot-action output within the same context.

Article

Reka releases 19B Rho-1: one model handles text and images, video, and robot actions

Beating AI news, the U.S.FAIR company Reka, founded by former researchers from DeepMind, Google Brain, and Meta,AI released a 19000000000 parameter model Rho-1. It can process text, images, and video in a single conversation, and can also output robot actions. For example, it can first generate an image of a lighthouse, then animate the lighthouse, change the weather in the video, and finally answer follow-up questions about the differences between two videos. Content generated earlier remains in the context and does not need to be handed over to another model for processing.

It also supports receiving instructions during video generation. While the video is being generated, users can temporarily request changes to the weather, direction of motion, or other content, and subsequent frames will be adjusted accordingly. The base version generates at approximately 0.79 times real time. The distilled version reduces the number of generation steps from 99 steps to 8 steps; Reka measured approximately 1 seconds to generate 5.3 seconds of video.

Robotics is also connected to the same model.Rho-1 predicts what the camera will see next while outputting the robot's action signals, without requiring a separate planning model to decide how to move. The official demonstrations currently consist of LIBERO simulation tasks, and no test results for physical robots have been released.

Similar unified models have also appeared this year. Nvidia Cosmos 3 can likewise process text, images, video, audio, and actions in a unified manner.Rho-1 focuses more on continuous interaction, emphasizing repeated generation, modification, and continued reasoning within the same context.Rho-1 was trained from scratch using 320 H100, and trained for approximately 3 months. The current video resolution is 672×384, and long videos may still exhibit structural drift; video localization and local editing are also not sufficiently stable. The model is currently only a research preview, and its weights and API have not been made publicly available.

Original article link https://m.theblockbeats.info/flash/370451

Key points

The article states that the U.S.AI company Reka released a 19000000000 parameter model Rho-1。
Rho-1 that can process text, images, and video in a single conversation and output robot actions.
The article states that instructions to change the weather, direction of motion, and other elements can continue to be received during video generation.
The base version generates at approximately 0.79 times real time; the distilled version compresses the generation steps from 99 steps to 8 steps, and Reka measured approximately 1 seconds to generate 5.3 seconds of video.
The article states that Rho-1 can simultaneously predict what the camera will see next and output robot action signals.
The current demonstrations consist of LIBERO simulation tasks, and no physical-robot test results have been released.
Rho-1 Using 320 H100 and training from scratch, it was trained for approximately 3 months.
The current video resolution is 672×384; long videos may exhibit structural drift, and video localization and local editing are also not sufficiently stable.
The model is currently a research preview, and its weights and API have not been made publicly available.
The article mentions that Nvidia Cosmos 3 can also process text, images, video, audio, and actions in a unified manner.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Rho-1 attempts to place multiple tasks into a single model and the same conversation context. After generating an image or video, users can continue modifying the content or asking about differences; in robotic tasks, the model can also simultaneously predict visual changes and provide action signals. However, the robot demonstrations disclosed so far are limited to simulation, and the model is not yet available for use.

Why it matters to readers

If the capabilities described in the article are subsequently validated,Rho-1 demonstrates a unified-model direction: reducing the need to switch models between text, image, video, and robotics tasks, while supporting continuous interaction during generation.

Beginners can understand it as a multifunctional AI model that remembers previously generated content, continues to accept modification requests, and attempts to control robot actions; however, it is still only a research preview and cannot be accessed through public weights or API directly.

Risks and unknowns

The materials provided are a media news brief and do not include an official Reka announcement or technical report for verification.
The speed, training resources, and model capability data cited in the article have not been supported by independent test results.
The robot capabilities have so far only been demonstrated in LIBERO simulation tasks; performance on physical robots is unknown.
Weights and API have not yet been released, so external users cannot reproduce the actual tests.

Reka releases 19B Rho-1 research preview for unified processing of text and images, video, and robot actions

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center