Reka AI launches Rho-1 omni-model handling text, images, video and robotics

Reka AI has released a research preview of Rho-1, a 19-billion-parameter omni-reasoning model that processes and generates text, images, video, and robot actions within a single neural network. Trained on 320 H100 GPUs over three months, Rho-1 eliminates multi-agent pipelines by running all modalities as tokens in one shared context window. The model generates real-time video capped at 672×384 and outputs continuous robot actions, utilizing an inverse dynamics model to learn from raw internet videos.

Release and availability of Rho-1

• ▪As of October 6, 2026, Rho-1 is available only as a research preview, with no public weights, API, or pricing released by Reka AI
• ▪On October 5, 2026, Reka AI released a research preview of Rho-1, a 19-billion-parameter omni-reasoning model that processes and generates text, images, video, and robot control actions in a single neural network

Architecture of the Rho-1 model

• ▪Reka AI's Rho-1 architecture uses two expert weight streams within each transformer block—an understanding stream for language and visual parsing, and a generation stream for denoising latents—which share attention and a single KV cache
• ▪Reka AI's Rho-1 training combines next-token prediction for discrete sequences, such as text and symbolic reasoning, with flow matching for continuous outputs, including image latents, video frames, and robot actions
• ▪Reka AI's Rho-1 processes all modalities as tokens inside a single shared context window, eliminating the need for tool calls, external models, or multi-agent pipelines

Video generation performance and specifications

• ▪A distilled variant of Reka AI's Rho-1 reduces denoising steps from 99 to 8, returning a 5.3-second video clip in approximately one second with minimal reported quality loss
• ▪Reka AI's base Rho-1 model generates video at 0.79x real-time median speed, delivering a watchable stream in approximately 6 seconds and a first clip in 7.0 seconds
• ▪The video generation capability of Reka AI's Rho-1 model is capped at a resolution of 672×384 pixels

Robotic control capabilities and training

• ▪To address scarce robot training data, Reka AI paired Rho-1 with an Inverse Dynamics Model that infers control signals from raw internet videos
• ▪Rho-1 outputs native continuous action tokens and future video frames from the same latent state, emitting 7 action channels in a LIBERO simulation episode

Alternative model features

• ▪Unlike ByteDance BAGEL, BAAI Emu3.5, and Google Genie 3 (which takes navigation inputs and supports steering), which do not output robot actions, Reka AI's Rho-1 emits native continuous action tokens and supports real-time steering
• ▪ByteDance BAGEL and BAAI Emu3.5 are distributed as open-weights models under the Apache 2.0 license, unlike Google Genie 3

Debatable claims

• ▪Single omni-models are more effective than multi-agent pipelines for multimodal AI
• ▪Developers should release the weights of advanced omni-reasoning models to the public