Reka previews Rho-1 as one model for text, media and physical actions
Reka published a research preview of Rho-1, a 19B model trained to understand and generate text, images and video and to emit action tokens within one network. The design interleaves discrete autoregressive decoding with continuous diffusion trajectories instead of delegating every modality to a separate specialist.
Why it matters: If the architecture transfers beyond demonstrations, one shared context could reduce pipeline handoffs and enable tighter visual or physical feedback loops. It also creates a harder serving and safety problem because generation, perception and action share…