Tencent AI Lab Releases HY-Embodied-0.5 Series Embodied AI Baseline Models

According to 1M AI News monitoring, Tencent HybridEx released the HY-Embodied-0.5 series Embodied AI base model, with the 2B parameter version officially open-sourced. This series of models is designed for real-world embodied intelligent agents, focusing on improving spatiotemporal perception and embodied reasoning abilities, and can be used in applications such as environment prediction, interactive execution, and task planning.

The model family includes two versions: 2B and 32B. The 2B model is aimed at edge devices and lightweight deployment, while the 32B model is used for complex reasoning tasks and high-level intelligent decision-making. In terms of technical architecture, HY-Embodied-0.5 introduces the Mixture-of-Transformers (MoT) mechanism to achieve multimodal division of labor computation, enhances perceptual representation with latent tokens, and combines self-evolving distillation and policy-based distillation methods to achieve collaborative optimization of model capacities.

In terms of performance, this series of models excels in 22 benchmark tests, with the 2B model outperforming state-of-the-art models of the same scale in 16 tasks, and the 32B version approaching the frontier level, demonstrating strong embodied intelligence generalization capabilities and practical application potential.