Tencent has released Hy-Embodied-RxBrain-1.0, a ~6.2B-parameter unified multimodal foundation model designed for embodied cognition that couples language reasoning with visual imagination.
- The model uses a Unified Mixture-of-Transformers (MoT) backbone where understanding and image synthesis share one autoregressive architecture.
- It features a flow-matching image head that decodes into a frozen FLUX VAE latent space to generate imagined frames.
- Key capabilities include embodied understanding, world state prediction, and joint subgoal planning through interleaved generation of text reasoning and visual goals.
This approach allows the model to decompose tasks into steps by emitting both the next action and the goal image it should reach.