According to a technical report, the Beijing Academy of Artificial Intelligence (BAAI) developed Orca, a world model that builds an internal picture of the world from image and language signals. Orca combines "unconscious learning" from raw videos without captions and "conscious learning" using verbal instructions, where videos are split into segments labeled with state changes. The pre-trained language-image model Qwen3.5 serves as a frozen core, with separate, swappable modules for text, images (via Stable Diffusion 3.5), and robot actions (via a module called "Action Expert"). The researchers trained Orca on 125,000 hours of video footage, 160 million event descriptions, and 11.5 million question-answer pairs, with only one-tenth of the video data used in the current version. The team argues that intelligence should be defined by models that build a general grasp of how the world changes, not by specialized prediction models.