Reka AI's Rho-1 Omni-Model Handles Text, Images, Video and Robot Control in One 19B Network

Reka AI has built a single neural network that reads and writes text, images and video, and also outputs robot control actions, rather than stitching together separate systems for each job. The model, Rho-1, carries 19 billion parameters and was trained on 320 Nvidia H100 GPUs across about three months, according to the company. That compute budget sits far below what frontier multimodal models typically consume, making the training recipe as notable as the capabilities themselves.

The architectural choice at the center of Rho-1 is that every input and output is treated as tokens inside one shared context window. Text, image patches, video frames and robot action commands all flow through the same attention mechanism, so the model can condition a motor command on a paragraph of instructions, a camera frame or both at once. Reka's approach contrasts with the dominant pattern in the field, where a language model routes requests to specialized vision, video or control models through an orchestration layer.

Robot control is the piece that separates Rho-1 from the current crop of omni-models. Most multimodal systems stop at understanding and generating media; adding action tokens means the same weights that caption an image can also drive a manipulator or a mobile base. That is a step toward general-purpose robot policies that do not need a bespoke control stack for every task or embodiment, though Reka has not yet published head-to-head benchmarks against dedicated robotics models.

The efficiency claim is the other headline number. Training a 19B multimodal model on 320 H100s for three months is a small fraction of the clusters used for leading open and closed models, which often run into the tens of thousands of GPUs. If Rho-1's results hold up, it suggests that careful data curation and unified tokenization can substitute for raw compute at this scale. That matters for labs and startups that cannot secure nine-figure hardware budgets.

Reka has not disclosed the full training dataset, the exact benchmark scores or whether Rho-1 will be released as open weights, an API or a product feature. Those details will determine whether the model is a research demonstration or something builders can actually use. The broader question is whether a single shared context genuinely outperforms modular pipelines on real tasks, or whether it mainly wins on simplicity and cost.

Expect the next round of scrutiny to focus on robot-control evaluations and on how Rho-1 handles long video inputs, where token counts explode and shared context becomes expensive. If Reka publishes weights or an accessible endpoint, independent replication on consumer-grade hardware will be the real test of the efficiency story. Until then, Rho-1 is a credible signal that omni-modal training is getting cheaper and that action tokens are joining the standard modality set.