Tencent-Hunyuan / HY-Embodied

What is HY-Embodied?

HY-Embodied delivers a family of Python-based embodied foundation models designed for real-world agents, tackling the gap between general vision-language models and the precise needs of physical robotics. You get pretrained weights—like the efficient HY-Embodied-0.5 MoT-2B for edge devices and a 32B powerhouse for heavy reasoning—that excel at spatial-temporal perception, prediction, interaction, and planning from images or video. Load them via Hugging Face Transformers for quick inference on tasks like robot control in Vision-Language-Action pipelines, outputting structured thoughts and normalized coordinates.

Why is it gaining traction?

It crushes competitors on 16+ embodied benchmarks, like 89.2 on CV-Bench and 82.8 on EmbSpatial-Bench, often with fewer active parameters for faster real-world deployment. Developers dig the plug-and-play inference scripts for single or batch runs, thinking mode for step-by-step reasoning, and seamless VLA integration without custom training. With 497 stars, it's pulling in robotics folks tired of underperforming VLMs in 3D spaces.

Who should use this?

Robotics engineers wiring up physical agents for manipulation or navigation tasks. AI researchers benchmarking embodied reasoning on spatial datasets. Hardware devs deploying lightweight vision models to edge robots handling real-world interactions like object grasping or trajectory planning.

Verdict

Grab it if you're in embodied AI—impressive benchmarks and ready inference make it a solid starting brain for agents, despite the 1.0% credibility score and modest 497 stars signaling early maturity. Docs are clear, but hold off for production until fine-tuning lands.

(198 words)