Reka has unveiled a research preview for Rho-1, a 19-billion parameter omni-reasoning model trained from scratch. The unified architecture handles text, images, video, and direct robotic control within a single neural network, bypassing traditional multi-agent pipelines. According to Reka, the model combines understanding and generation into a single context window using twin expert streams inside each transformer block.

In internal testing, the model successfully generated visual media, edited content, and outputted robotic movement instructions across 7 action channels without calling external tools or modular models. Reka implemented flow matching for continuous outputs alongside standard next-token prediction, allowing the model to produce video at 0.79x real-time speed. A distilled variant reduced denoising steps from 99 to 8, generating video clips in roughly one second.

To overcome limits in physical robotic training data, Reka paired Rho-1 with an Inverse Dynamics Model designed to extract control signals directly from video sources. Reka position the unified model as a faster alternative to multi-model agentic pipelines, citing reduced latency and shared key-value cache efficiency.

Why it matters

  • Developers can reduce latency by replacing complex multi-model pipelines with unified omni-reasoning networks.

  • Robotics teams can leverage raw video data using inverse dynamics models to overcome teleoperation data shortages.

  • AI architects gain efficiency by sharing context across generation and understanding streams inside single transformers.

Source: marktechpost.com