Reka has put video, images, text and robot control into one 19-billion-parameter model
The same weights that predict the next video frame also produce the next robot action. Reka says it trained the model on 320 H100 GPUs in roughly three months, and is releasing it as a research preview rather than open weights.