Tencent Hunyuan releases the HY-Embodied-0.5 series of embodied intelligence models, and the 2B version is now open-sourced

Gate News message, April 10: Tencent Hunyuan released the HY-Embodied-0.5 series embodied intelligence foundation models, and the 2B-parameter version has been officially open-sourced. The series models are designed for embodied intelligence agents in the real world, focusing on enhancing spatiotemporal perception and embodied reasoning capabilities, and can be applied to scenarios such as environmental prediction, interaction execution, and task planning.

The model family includes two versions: 2B and 32B. The 2B model is intended for edge devices and lightweight deployment, while the 32B model is used for complex reasoning tasks and higher-level intelligent decision-making. In terms of technical architecture, HY-Embodied-0.5 introduces a Mixture-of-Transformers (MoT) mechanism to implement multimodal partitioned computation, enhances perceptual representation capabilities through latent Tokens, and combines self-evolution after-training with policy-based distillation to achieve coordinated capability optimization across small and large models.

In terms of performance, the series models have performed exceptionally well across 22 benchmark tests. The 2B model surpasses same-scale SOTA models on 16 tasks, and the 32B version is already close to frontier levels, demonstrating strong embodied intelligence generalization capability and real-world application potential.