Tencent did not lead with a robot dog, arm or booth demo. It put two “brains” into the open-source community: Hy-Embodied-VLM-1.0 for understanding real scenes and reasoning about action effects, and Hy-Embodied-RxBrain-1.0 for breaking tasks into language steps while imagining the target image after each step.
The point is that robot systems often split seeing, thinking and acting across separate stacks. Tencent is pulling two upper cognitive layers toward foundation-model form: one model pursues deployment efficiency with about 3B activated parameters, while the other uses about 6.2B parameters to combine reasoning and visual goal generation.
Two models, two jobs
Hy-Embodied-VLM-1.0 is a second-generation embodied VLM. It uses the Hy3-A3B language backbone and Hy-ViT2 vision encoder, with about 30B total MoE parameters and around 3B activated per token. Tencent says it ranked first among same-size models on 19 of 38 embodied benchmarks and second on another 11, improving the previous Hy-Embodied-0.5 MoT-2B average by 8.4%.
“We have released Hy-Embodied-VLM-1.0!”
RxBrain is more about imagination and planning. Tencent defines it as an embodied cognition foundation model for image and multi-frame video QA, short-horizon world-state prediction and joint subgoal planning. It alternates reasoning text and imagined images in one autoregressive sequence, using a learned <Image> token to decide when to generate a target frame.
Fast brain and slow brain
- VLM-1.0 focuses on efficiency: about 30B total parameters, about 3B active parameters, BF16 weights around 86GB, and recommended H100, H20 or A100 80GB environments.
- RxBrain focuses on planning expression: about 6.2B parameters for text, image, video and interleaved image-text generation.
- Both are open source, with repositories and licenses pointing to Apache 2.0.
The split resembles a fast brain and a slow brain. The fast side must be cheap and stable enough for repeated deployment calls. The slow side breaks long tasks down and turns abstract language into visual goals that can be checked.
The data account matters
VLM-1.0 used more than 18 million QA samples in mid-training and about 48,000 high-quality instruction samples in post-training. RxBrain's data budget is heavier: 50,177 hours of operation data, including 31,568 hours of egocentric and UMI data, 17,292 hours of real-robot data and 1,317 hours of simulation data. Open-source data accounted for 28,597 hours, about 57%.
The team further cut unlabeled long videos into atomic action sequences with action descriptions, start-end states and time boundaries, forming about 210 million training samples. Mid-training added 35 million embodied-capability samples covering spatial reasoning, multi-view understanding, causal inference, visual grounding and behavior planning.
Real-robot results still show limits
On RxBrain-Bench, joint embodied planning scored 0.68, above Cosmos3-Nano Agent at 0.521, BAGEL-7B-MoT at 0.503 and modular Qwen-Agent at 0.431. For short-horizon future video generation, RxBrain scored 0.62, above Wan2.2-TI2V-5B at 0.429 and close to Cosmos3-Nano at 0.591.
Real-robot validation used DOBOT X-Trainer and ARX A5 arms on tableware placement, folding and storing glasses, and throwing trash away. Reported success rates were 97%, 95% and 68%, averaging 87%; π0 and π0.5 baselines averaged 68% and 82%. The weak points are also visible: target-image correctness scored 0.52, temporal and physical plausibility 0.53, and planning fell from 0.69 at two steps to 0.55 at eight steps.
Open source adds a benchmark
The meaningful contribution is not that Tencent also has robot models. By publishing VLM, world-state prediction, interleaved planning, weights and inference code, it gives China's embodied-AI ecosystem a reproducible reference. The next test is whether developers can download weights, run inference through vLLM and Hugging Face transformers, and transfer real tasks beyond Tencent's data and paper settings.
Sources: Synced; arXiv:2607.12894 was used to verify model parameters, 38 evaluations and performance framing; GitHub and Hugging Face verified open weights, Apache 2.0 license and hardware requirements; CocoLoop; Tencent Hunyuan public repositories verified RxBrain parameters, data scale, benchmark scores and real-robot success rates.