Blog Network

Tencent Robotics X & HY Vision · 2026-04-09 · major

HY-Embodied-0.5

Tencent ships open-weights embodied foundation models in two sizes (2B MoT and 32B MoE) — tuned for spatial-temporal perception and robot VLA control, beating same-size peers on 16 of 22 benchmarks.

HY-Embodied GitHub repository social card

Open-weights embodied foundation models from Tencent — built to give robot VLA stacks a real spatial-temporal brain.

Key specs

VariantsMoT-2B (2.2B active) / MoE-A32B
Training tokens200B+
Benchmarks beaten16 of 22

What is it?

HY-Embodied-0.5 is a family of open-weights foundation models from Tencent Robotics X and the Hunyuan vision team, aimed specifically at physical agents. It ships in two sizes: a 2B Mixture-of-Transformers model with 2.2B activated parameters for edge deployment, and a larger 32B MoE model for heavier reasoning. The weights, paper and inference code for the 2B variant were open-sourced on April 9.

How does it work?

The architecture is a Mixture-of-Transformers that uses latent tokens to route computation per modality (vision, language, action) instead of a single dense stack. The 2B model is distilled on-policy from the 32B teacher. Training uses over 100 million embodied and spatial data points and more than 200 billion tokens, with special attention to spatial-temporal perception. The models are designed to plug into Vision-Language-Action pipelines as the perception and reasoning 'brain' that drives a downstream robot controller.

Why does it matter?

Most general-purpose VLMs stumble on the exact things embodied agents need — depth, motion, object persistence, plan consistency. HY-Embodied-0.5 is the first Tencent-backed open release purpose-built for that gap, and it beats similarly-sized models on 16 of 22 embodied benchmarks while the 32B variant tracks frontier models like Gemini 3.0 Pro. For robotics teams that cannot ship a closed frontier API on a real robot, this is a usable foundation.

Who is it for?

Robotics / embodied-AI researchers, VLA framework authors.

Try it

huggingface.co/tencent/HY-Embodied-0.5

Sources · 3 outlets

Tags

  • embodied-ai
  • vla
  • open-weights
  • robotics
  • multimodal

← All releases