What is Hy-Embodied-VLM-1.0 and what are its key performance figures?

Hy-Embodied-VLM-1.0 is a new embodied AI foundation model built on the Hy3-A3B language backbone and Hy-ViT2 vision encoder with a Mixture-of-Experts architecture. It achieves state-of-the-art on 19 of 38 benchmarks, activates only 3B parameters while matching 32B models, and improves 8.4% over its predecessor.

Why does this model matter for embodied AI and robotics?

By delivering 32B-level performance with only 3B active parameters, it dramatically reduces computational overhead and enables deployment on edge devices. Its action-centric capability framework opens new paths for robot control and latency-sensitive applications.

What are the next steps and future research directions?

Ablation studies show transformational reasoning and sequence-adaptive reasoning are critical for complex long-horizon tasks. Future work will focus on optimizing these dimensions to improve real-world physical interaction and multi-turn decision-making.

Hy-Embodied-VLM-1.0：建構高效物理世界具身智能基礎模型

本文介紹了Hy-Embodied-VLM-1.0，一款專為物理世界交互設計的高效具身基礎模型。針對具身智能體在感知、推理及動態適應方面的挑戰，研究提出了一種以動作為核心的能力分類體系，涵蓋狀態理解、轉換推理及序列自適應推理三個維度。模型基於Hy3-A3B語言骨干和Hy-ViT2視覺編碼器構建，採用混合專家架構以平衡容量與推理效率。在涵蓋38個基準的廣泛評估中，該模型在19個任務上達到同類最佳水平，顯著優於Qwen3.6-A3B等競品，相比上一代模型平均性能提升8.4%，且僅激活3B參數即可媲美32B參數模型，展現了卓越的推理與交互能力。

Sources

arXiv