Tencent RxBrain (Hy-Embodied-RxBrain-1.0)

RxBrain is Tencent's attempt to give a robot an imagination in the literal sense: alongside the sentence describing the next step, the model generates the image of the world after that step. Tencent Robotics X published the weights and technical report in July 2026 under Apache 2.0. Most planning models emit text. RxBrain emits both, interleaved in a single sequence — a learned image token decides when to stop writing and start picturing. Each planned subgoal therefore comes as a pair: the action in words and the target frame in pixels. The stated purpose is to check a plan against a visual expectation rather than against a description that may be satisfied in the wrong way. The backbone is a mixture-of-transformers of about 6.2 billion parameters, with separate pathways for text, vision and image generation inside one autoregressive model instead of separate towers. Imagined frames come from a flow-matching head decoding into the latent space of a frozen FLUX autoencoder — Tencent did not train its own image decoder, it borrowed one and kept it fixed. Three capabilities are claimed: answering questions and reasoning over images and multi-frame video, predicting the near-future frames an action would produce, and decomposing a task into steps that carry both an instruction and a goal image. The honest caveat is the same as for the rest of this family: the results are the vendor's own, published in arXiv 2607.14187, and we found no independent evaluation. What is checkable is that the weights, the inference code and a public demonstration space are all available, and the licence is plain Apache 2.0 with no territorial exclusion — unlike the sibling HY-Embodied-0.5-X, which does not apply in the European Union. Over the 30 days to 31 August 2026 the model drew 149 downloads. That is a small number, and it should be read as a research release rather than a deployed product.

▸Videos

No videos yet.

▸Reviews