Tencent Robotics X and the HY Vision Team released HY-Embodied-0.5, a suite of open-source foundation models built for embodied agents in real-world physical environments, on April 8. The family uses a Mixture-of-Transformers architecture with two variants: a 2B model for edge deployment and a 32B MoE model for complex reasoning. The 2B model achieved best results in 16 of 22 benchmarks spanning perception, reasoning, and planning, outperforming similarly-sized competitors such as Qwen3-VL-4B; the 32B variant matches Gemini 3.0 Pro on comprehensive evaluations. Weights and inference code are publicly available on Hugging Face and GitHub.