Tencent Robotics X, Futian Laboratory, and Tencent Hunyuan have released the technical report, inference code, and weights for Hy-Embodied-RxBrain-1.0, a ~6.2B-parameter unified multimodal foundation model for embodied cognition.
The model couples language reasoning with visual imagination through interleaved autoregressive generation, alternating text and flow-matched imagined frames within a single sequence. Key architectural components include a Unified Mixture-of-Transformers backbone with modality-specific pathways and a Flow-Matching Image Head that decodes into a frozen FLUX VAE latent space.
RxBrain delivers three core capabilities: embodied understanding and reasoning over images and video, world state prediction to imagine near-future frames, and joint subgoal planning that decomposes tasks into steps with both language actions and goal images.