Reka AI releases Rho-1, a 19B-parameter omni-model for text, image, video and robot control

Reka AI has released a research preview of Rho-1, a 19-billion-parameter omni-model that processes and generates text, images, video, and robot control actions within a single neural network, according to The Decoder's report on the release.

What Rho-1 does

Most multimodal systems route tasks to specialized models or call external tools. Rho-1 does neither, according to Reka. All modalities run as tokens in one shared context window, with no tool calls and no external models.

Reported capabilities:

• Continuous video generation in real time. The model generates video as a continuous stream.
• Live instruction updates. It responds to new instructions on the fly without restarting.
• Unified perception and control. The same weights that predict camera images also drive robot movements.

Training approach

Robot training data is scarce. To work around this, Reka built an inverse dynamics model that extracts control signals from ordinary internet videos. That lets the team use large volumes of unlabeled video as a source of action data.

Rho-1 was trained on 320 H100 GPUs over about three months. Assuming continuous use, that implies roughly 700,000 H100 GPU-hours. This is our estimate, not a figure Reka has published.

What has not been disclosed

The source material does not include:

• Context window size: not yet disclosed
• Pricing per 1M tokens: not yet disclosed
• Benchmark scores: none published in the source report
• Training data cutoff: not yet disclosed
• Weights, license, and API availability: not specified

The capability descriptions above, including real-time video generation and unified robot control, come from Reka. We have not seen independent evaluation.

Background

Reka AI is not new to multimodal work. In April 2024 it shipped Reka Core, a multimodal language model that, per the company at the time, competed with GPT-4, Claude 3, and Gemini Ultra on benchmarks. Rho-1 extends the company's focus from understanding multimodal inputs to generating video and actions.

The release fits a broader push in AI research toward so-called world models, which learn to predict and act within visual environments rather than only produce text.

What this means

The notable claim is architectural: one set of weights handling perception, video prediction, and motor control as tokens in a single context. If it holds up, it removes the orchestration layer that most robotics stacks need between a vision-language model and a low-level controller.

Two caveats apply. First, this is a research preview with no published benchmarks, so there is no way yet to compare Rho-1 against specialized robot policies or dedicated video generators. Second, a 19B-parameter model generating video in real time raises questions about latency, resolution, and hardware requirements that Reka has not answered publicly.

The inverse dynamics approach to mining internet video for control signals may prove the more durable contribution. It addresses the main bottleneck in robot learning, the shortage of action-labeled data, without large-scale teleoperation collection. Its quality depends on how accurately the inferred actions match real robot behavior, which independent testing will need to confirm.

Until Reka releases technical details, access terms, and evaluations, Rho-1 is best treated as a research signal rather than a deployable option.